# Comparison: what Niadra has and what each memory system documents

> Identity across channels, memory shared across vendors, CRM and ERP events, state with dates, the claim check, coordination, turn records, permissions, receipts, erasure, prompt-ready context, voice, MCP, benchmark and price: Niadra against Mem0, Zep, Letta, Supermemory, Hindsight, Honcho, Cognee, AgentCore Memory and Twilio, with the source and the date of every cell.

URL: https://niadra.com/en/comparativo

# What each one has. With the source of every cell.

Sixteen capabilities, Niadra and ten memory systems. Every cell says what the vendor’s own documentation states, with the link to the page and the day it was read. Where a competitor has what Niadra does not, the row says so. For the measured numbers, see the benchmark.

-   yesthe vendor’s documentation describes the whole capability
-   partialit describes part of it; the note says which
-   nothe documentation says it works another way
-   not documentedthe page read says nothing about it, which does not mean it does not exist

Every mark is a link to the page read, on 2026-10-01. Documentation changes: a cell only changes after a new reading, with a new date.

Capability

Niadra

Mem0 (open source)

Mem0 Platform

Zep

Letta

Supermemory

Hindsight

Honcho

Cognee

Amazon Bedrock AgentCore Memory

![](/integrations/logos/twilio.png)Twilio Conversation Memory

Identity across channelsThe memory works out that the phone, the WhatsApp number, the e-mail and each system’s ID belong to the same person. The application does not hand over a ready-made ID.

[yes](https://docs.niadra.com/concepts/identity "Niadra: yes. Identifiers by type, with a weight, reversible merges and a verification level per conversation. (read on 2026-10-01)")

[no](https://github.com/mem0ai/mem0 "Mem0 (open source): no. Memory is kept under the user_id the application sends. (read on 2026-10-01)")

[no](https://docs.mem0.ai/platform/features/entity-scoped-memory "Mem0 Platform: no. user_id, agent_id, app_id and run_id come from the application. (read on 2026-10-01)")

[no](https://help.getzep.com/users-and-user-graphs "Zep: no. The user_id \"maps to your internal User ID\". (read on 2026-10-01)")

[no](https://docs.letta.com/api-reference/identities/create "Letta: no. An identity carries the key the application chooses (identifier_key). (read on 2026-10-01)")

[no](https://supermemory.ai/docs/concepts/user-profiles "Supermemory: no. One profile per containerTag, set by the application. (read on 2026-10-01)")

[no](https://docs.hindsight.vectorize.io/memory-banks "Hindsight: no. One memory bank per user, project or agent, isolated from the others. (read on 2026-10-01)")

[no](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: no. Each peer carries the identifier the application gives it. (read on 2026-10-01)")

[no](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview "Cognee: no. Users and tenants come from the application; memory is per dataset. (read on 2026-10-01)")

[no](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html "Amazon Bedrock AgentCore Memory: no. The actorId is set by the application on every event. (read on 2026-10-01)")

[yes](https://www.twilio.com/docs/conversations/memory/identity-resolution "Twilio Conversation Memory: yes. Phone, e-mail, WhatsApp and user ID become one profile, only for conversations that flow through Twilio’s channels. (read on 2026-10-01)")

One memory for agents from different vendorsAgents from different vendors read and feed the same customer memory, each with its own permission, and cutting one vendor off erases nothing.

[yes](https://docs.niadra.com/integrations/overview "Niadra: yes. Each vendor has its own credential and permission; none reads the other’s memory or numbers. (read on 2026-10-01)")

[partial](https://github.com/mem0ai/mem0 "Mem0 (open source): partial. Any agent with the key reads the same user_id. No permission per vendor. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/platform/features/entity-scoped-memory "Mem0 Platform: partial. Several agents filter the same user_id by agent_id. No permission per vendor. (read on 2026-10-01)")

[partial](https://help.getzep.com/memory-mcp-server "Zep: partial. The company’s agent and the person’s MCP client share one user graph. (read on 2026-10-01)")

[partial](https://docs.letta.com/concepts/shared-memory "Letta: partial. Blocks shared between Letta agents, through git repositories. Other vendors are not mentioned. (read on 2026-10-01)")

[partial](https://supermemory.ai/docs/authentication "Supermemory: partial. Keys restricted to containerTags; any agent with the key reads the profile. (read on 2026-10-01)")

[partial](https://docs.hindsight.vectorize.io/api-keys "Hindsight: partial. Keys restricted to banks; any agent with the key reads the bank. (read on 2026-10-01)")

[partial](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: partial. Several peers in a session, with a setting for who observes whom. (read on 2026-10-01)")

[partial](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview "Cognee: partial. Datasets shared through an access list (read, write, delete, share). (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-cross-account-access.html "Amazon Bedrock AgentCore Memory: partial. Any IAM principal, even from another AWS account, reads the same memory. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/conversations/memory/api-quickstart "Twilio Conversation Memory: partial. Any agent calls the recall API with the account key; the conversation must flow through Twilio. (read on 2026-10-01)")

Events from the CRM and the ERPOrders, bills and tickets enter the memory through the events the system already emits, tied to the right customer by that system’s ID.

[yes](https://docs.niadra.com/concepts/systems "Niadra: yes. Webhook with a versioned mapping, batch and file; no AI in the path. (read on 2026-10-01)")

[not documented](https://docs.mem0.ai/open-source/overview "Mem0 (open source): not documented (read on 2026-10-01)")

[not documented](https://docs.mem0.ai/platform/features/webhooks "Mem0 Platform: not documented. Mem0’s webhooks are outbound: they notify when a memory changes. (read on 2026-10-01)")

[partial](https://help.getzep.com/adding-business-data "Zep: partial. Business records go in as JSON through graph.add; no inbound webhook and no per-system ID. (read on 2026-10-01)")

[not documented](https://docs.letta.com/agent-sdk/memory "Letta: not documented (read on 2026-10-01)")

[not documented](https://supermemory.ai/docs/connectors/google-drive "Supermemory: not documented. The connectors are knowledge sources: Drive, Gmail, Notion, OneDrive. (read on 2026-10-01)")

[not documented](https://docs.hindsight.vectorize.io/retain "Hindsight: not documented (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/v3/documentation/features/advanced/webhooks "Honcho: not documented. Honcho’s webhooks are outbound: they notify when the queue empties. (read on 2026-10-01)")

[not documented](https://www.cognee.ai/pricing "Cognee: not documented. The data sources are Slack, Notion, Linear and Google Drive. (read on 2026-10-01)")

[not documented](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-terminology.html "Amazon Bedrock AgentCore Memory: not documented. An event is a message, a system action or a tool call of the conversation. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/conversations/memory/api-quickstart "Twilio Conversation Memory: partial. Profiles and traits go in through bulk import; no order or bill event. (read on 2026-10-01)")

Typed business state with freshnessAn order, a bill, a quote: every value reaches the agent with its type, the date it was last updated, the source and what may not be stated with it.

[yes](https://docs.niadra.com/concepts/object-types "Niadra: yes. Types declared or derived from the database; freshness per field computed on the read. (read on 2026-10-01)")

[no](https://github.com/mem0ai/mem0 "Mem0 (open source): no. Memory is text extracted from the conversation. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/platform/features/timestamp "Mem0 Platform: partial. A creation date and an expiry per memory; no typed object and no freshness per field. (read on 2026-10-01)")

[partial](https://help.getzep.com/facts "Zep: partial. Facts with valid_at and invalid_at; no typed object per field. (read on 2026-10-01)")

[no](https://docs.letta.com/agent-sdk/memory "Letta: no. Memory blocks in text, edited by the agent. (read on 2026-10-01)")

[partial](https://supermemory.ai/docs/concepts/graph-memory "Supermemory: partial. Facts with update relations and expiry; no typed object per field. (read on 2026-10-01)")

[no](https://docs.hindsight.vectorize.io/retain "Hindsight: no. Text items, with an optional timestamp. (read on 2026-10-01)")

[no](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: no. Representations and conclusions about the person, in text. (read on 2026-10-01)")

[partial](https://docs.cognee.ai/guides/time-awareness "Cognee: partial. Events with timestamps and fact validity; no business object per field. (read on 2026-10-01)")

[no](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html "Amazon Bedrock AgentCore Memory: no. Facts, preferences and summaries extracted from the conversation. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: partial. Factual memory as key-value traits; no freshness per field. (read on 2026-10-01)")

The claim checkBefore the answer goes out, every price, date and number the agent is about to say is checked against what it consulted. What has no source is flagged.

[yes](https://docs.niadra.com/concepts/claims "Niadra: yes. A contract per category; the check runs in the agent’s process, with no extra AI. (read on 2026-10-01)")

[not documented](https://github.com/mem0ai/mem0 "Mem0 (open source): not documented (read on 2026-10-01)")

[not documented](https://docs.mem0.ai/platform/overview "Mem0 Platform: not documented (read on 2026-10-01)")

[not documented](https://help.getzep.com/concepts "Zep: not documented (read on 2026-10-01)")

[not documented](https://docs.letta.com/agent-sdk/memory "Letta: not documented (read on 2026-10-01)")

[not documented](https://supermemory.ai/docs/concepts/graph-memory "Supermemory: not documented (read on 2026-10-01)")

[not documented](https://docs.hindsight.vectorize.io/retain "Hindsight: not documented (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: not documented (read on 2026-10-01)")

[not documented](https://github.com/topoteretes/cognee "Cognee: not documented (read on 2026-10-01)")

[not documented](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html "Amazon Bedrock AgentCore Memory: not documented (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: not documented (read on 2026-10-01)")

Coordination between agentsBefore acting, the agent asks who holds the customer and what was already done. One business fact causes one effect.

[yes](https://docs.niadra.com/concepts/coordination "Niadra: yes. Ownership by declaration and observation, effect exactly once, a contact token for the gateway. (read on 2026-10-01)")

[not documented](https://github.com/mem0ai/mem0 "Mem0 (open source): not documented (read on 2026-10-01)")

[not documented](https://docs.mem0.ai/platform/overview "Mem0 Platform: not documented (read on 2026-10-01)")

[not documented](https://help.getzep.com/concepts "Zep: not documented (read on 2026-10-01)")

[not documented](https://docs.letta.com/concepts/shared-memory "Letta: not documented. The shared memory page says nothing about locks or ownership. (read on 2026-10-01)")

[not documented](https://supermemory.ai/docs/concepts/graph-memory "Supermemory: not documented (read on 2026-10-01)")

[not documented](https://docs.hindsight.vectorize.io/memory-banks "Hindsight: not documented (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: not documented (read on 2026-10-01)")

[not documented](https://github.com/topoteretes/cognee "Cognee: not documented (read on 2026-10-01)")

[not documented](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html "Amazon Bedrock AgentCore Memory: not documented (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: not documented (read on 2026-10-01)")

Turn records and replayWhat the agent read, called, showed and stated on every turn, with the build pinned, and the replay of a real conversation in your CI.

[yes](https://docs.niadra.com/concepts/turn-records "Niadra: yes. Captured in the agent’s process; replay in CI with the tools answering from the record. (read on 2026-10-01)")

[not documented](https://github.com/mem0ai/mem0 "Mem0 (open source): not documented (read on 2026-10-01)")

[not documented](https://docs.mem0.ai/platform/overview "Mem0 Platform: not documented (read on 2026-10-01)")

[not documented](https://help.getzep.com/concepts "Zep: not documented (read on 2026-10-01)")

[partial](https://docs.letta.com/api-reference/steps/list "Letta: partial. The steps API keeps the model, tokens and trace id of each step; replay is not documented. (read on 2026-10-01)")

[not documented](https://supermemory.ai/docs/concepts/graph-memory "Supermemory: not documented (read on 2026-10-01)")

[not documented](https://docs.hindsight.vectorize.io/retain "Hindsight: not documented (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: not documented (read on 2026-10-01)")

[not documented](https://github.com/topoteretes/cognee "Cognee: not documented (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-observability.html "Amazon Bedrock AgentCore Memory: partial. Extraction metrics and logs in CloudWatch; replay is not documented. (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: not documented (read on 2026-10-01)")

Per-agent permissions and field accessEach agent reads only what its purpose allows, item by item, and a sensitive field reaches only who may see it.

[yes](https://docs.niadra.com/concepts/privacy "Niadra: yes. Denied by default, granted by purpose; fixed sensitive categories; field access on the state. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/open-source/features/rest-api "Mem0 (open source): partial. An API key per server user; no purpose and no field. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/api-reference/organizations-projects "Mem0 Platform: partial. Reader and owner roles per project; no purpose and no field. (read on 2026-10-01)")

[partial](https://help.getzep.com/governance "Zep: partial. Policies on API keys and user groups; no purpose and no field. (read on 2026-10-01)")

[partial](https://docs.letta.com/pricing "Letta: partial. Role-based access control on the Enterprise plan. (read on 2026-10-01)")

[partial](https://supermemory.ai/docs/authentication "Supermemory: partial. Keys restricted to containerTags; no purpose and no field. (read on 2026-10-01)")

[partial](https://docs.hindsight.vectorize.io/api-keys "Hindsight: partial. Keys restricted to banks; no purpose and no field. (read on 2026-10-01)")

[partial](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: partial. A setting for which peers observe which; no purpose and no field. (read on 2026-10-01)")

[partial](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview "Cognee: partial. Read, write, delete and share per dataset; no agent and no field. (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html "Amazon Bedrock AgentCore Memory: partial. IAM policies by namespace and actor; no field access. (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory/api-quickstart "Twilio Conversation Memory: not documented. Authentication with the account key; no permission per agent. (read on 2026-10-01)")

Receipts of who read whatEvery read, the refused one included, leaves a record of which agent read what, for what and when, tamper-evident.

[yes](https://docs.niadra.com/concepts/receipts "Niadra: yes. A receipt per read, chained by SHA-256, with a daily root in write-once storage. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/open-source/features/rest-api "Mem0 (open source): partial. A recent log of the server’s API calls. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/platform/platform-vs-oss "Mem0 Platform: partial. A project event feed (add, search, delete); audit logs on Enterprise. (read on 2026-10-01)")

[partial](https://help.getzep.com/governance "Zep: partial. Audit logs of API and dashboard activity; no receipt per read. (read on 2026-10-01)")

[not documented](https://docs.letta.com/pricing "Letta: not documented (read on 2026-10-01)")

[not documented](https://supermemory.ai/docs/authentication "Supermemory: not documented (read on 2026-10-01)")

[not documented](https://docs.hindsight.vectorize.io/security/webhooks-siem "Hindsight: not documented. The SIEM receives only Memory Defense violations, with the key that wrote. (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/v3/documentation/features/advanced/deleting-data "Honcho: not documented (read on 2026-10-01)")

[not documented](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview "Cognee: not documented (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-cross-account-access.html "Amazon Bedrock AgentCore Memory: partial. CloudTrail logs the API calls; no chained receipt per read. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/conversations/memory/observability-logs "Twilio Conversation Memory: partial. Error and warning logs of the operations, retrievals included. (read on 2026-10-01)")

Erasure with a receiptErasing a customer walks through everything derived from their data and returns a receipt of what was erased.

[yes](https://docs.niadra.com/concepts/privacy "Niadra: yes. Erasure by lineage, with a receipt and the list of what to erase in your exported copy. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/open-source/features/rest-api "Mem0 (open source): partial. Deletion per memory or user; no receipt. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/api-reference/entities/delete-user "Mem0 Platform: partial. Deleting a user is asynchronous and returns an event_id; no receipt. (read on 2026-10-01)")

[partial](https://help.getzep.com/sdk-reference/user/delete "Zep: partial. Deletes the user, the threads and the graph; returns a success message. (read on 2026-10-01)")

[partial](https://docs.letta.com/api-reference/agents/delete "Letta: partial. Deleting an agent returns an empty object. (read on 2026-10-01)")

[partial](https://supermemory.ai/docs/api-reference/content-management/forget-a-memory "Supermemory: partial. Forgetting is a soft delete; it returns the id and the forgotten flag. (read on 2026-10-01)")

[yes](https://docs.hindsight.vectorize.io/security/data-privacy "Hindsight: yes. Two-step erasure with a receipt in the privacy events ledger; Enterprise only. (read on 2026-10-01)")

[partial](https://honcho.dev/docs/v3/documentation/features/advanced/deleting-data "Honcho: partial. Asynchronous deletion per session; derived conclusions survive; no receipt. (read on 2026-10-01)")

[partial](https://docs.cognee.ai/core-concepts/main-operations/forget "Cognee: partial. forget returns ids and a status; no receipt. (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/APIReference/API_DeleteMemoryRecord.html "Amazon Bedrock AgentCore Memory: partial. Deleting a record returns its id; no receipt. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/api/memory/v1/profile/delete-profile "Twilio Conversation Memory: partial. Deleting a profile is accepted (202) and irreversible; no receipt. (read on 2026-10-01)")

Prompt-ready compact contextThe agent receives text ready for the prompt, assembled before the request, with no AI on the read.

[yes](https://docs.niadra.com/concepts/context "Niadra: yes. Compiled when the memory changes, per agent and task, in layers the provider reuses. (read on 2026-10-01)")

[partial](https://github.com/mem0ai/mem0 "Mem0 (open source): partial. Search returns a list of memories; the application builds the prompt. (read on 2026-10-01)")

[partial](https://docs.mem0.ai/platform/overview "Mem0 Platform: partial. Search returns a list of memories; the application builds the prompt. (read on 2026-10-01)")

[yes](https://help.getzep.com/retrieving-context "Zep: yes. A context block assembled as text; ingestion can take minutes. (read on 2026-10-01)")

[partial](https://docs.letta.com/agent-sdk/memory "Letta: partial. The blocks live in the Letta agent’s prompt; they are not served to another agent. (read on 2026-10-01)")

[yes](https://supermemory.ai/docs/concepts/user-profiles "Supermemory: yes. A static and a dynamic profile, for the prompt. (read on 2026-10-01)")

[partial](https://docs.hindsight.vectorize.io/recall "Hindsight: partial. recall returns memories; reflect returns an answer a model writes on the read. (read on 2026-10-01)")

[yes](https://honcho.dev/docs/v3/documentation/core-concepts/architecture "Honcho: yes. get_context assembles conclusions and recent messages for the prompt. (read on 2026-10-01)")

[partial](https://github.com/topoteretes/cognee "Cognee: partial. Search answers with a model at read time. (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html "Amazon Bedrock AgentCore Memory: partial. Retrieval returns records; the application builds the prompt. (read on 2026-10-01)")

[partial](https://www.twilio.com/docs/conversations/memory/api-quickstart "Twilio Conversation Memory: partial. Recall returns structured observations and summaries; the application builds the prompt. (read on 2026-10-01)")

Voice integrationsA documented adapter for voice platforms: LiveKit, Pipecat, Vapi, Retell, ElevenLabs, Twilio.

[yes](https://docs.niadra.com/integrations/overview "Niadra: yes. LiveKit, Pipecat, Vapi, Retell, ElevenLabs and Twilio, with the context before the hello. (read on 2026-10-01)")

[yes](https://docs.mem0.ai/integrations/livekit "Mem0 (open source): yes. LiveKit, ElevenLabs and Pipecat. (read on 2026-10-01)")

[yes](https://docs.mem0.ai/integrations/livekit "Mem0 Platform: yes. LiveKit, ElevenLabs and Pipecat. (read on 2026-10-01)")

[yes](https://help.getzep.com/memory-for-agent-frameworks "Zep: yes. ElevenLabs and LiveKit. (read on 2026-10-01)")

[yes](https://docs.letta.com/guides/voice/livekit/ "Letta: yes. LiveKit and Vapi, marked experimental. (read on 2026-10-01)")

[yes](https://supermemory.ai/docs/integrations/livekit "Supermemory: yes. LiveKit and Pipecat. (read on 2026-10-01)")

[yes](https://github.com/vectorize-io/hindsight "Hindsight: yes. Pipecat, Vapi and ElevenLabs. (read on 2026-10-01)")

[not documented](https://honcho.dev/docs/llms.txt "Honcho: not documented. The documented integrations are LangGraph, CrewAI, Vercel AI SDK and n8n. (read on 2026-10-01)")

[not documented](https://github.com/topoteretes/cognee "Cognee: not documented. The documented integrations are coding assistants and MCP. (read on 2026-10-01)")

[not documented](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html "Amazon Bedrock AgentCore Memory: not documented (read on 2026-10-01)")

[yes](https://www.twilio.com/en-us/blog/developers/tutorials/product/orchestrate-multi-call-conversations-with-llm-twilio-conversation-memory "Twilio Conversation Memory: yes. Voice through Twilio itself, with ConversationRelay. (read on 2026-10-01)")

MCP serverThe memory as MCP tools, for any agent that speaks the protocol.

[yes](https://docs.niadra.com/guides/mcp "Niadra: yes. Eight tools, with the customer bound out of the model’s reach. (read on 2026-10-01)")

[yes](https://mem0.ai/blog/introducing-openmemory-mcp "Mem0 (open source): yes. OpenMemory MCP, local, on top of open-source Mem0. (read on 2026-10-01)")

[yes](https://docs.mem0.ai/platform/mem0-mcp "Mem0 Platform: yes. A hosted server with eleven tools. (read on 2026-10-01)")

[yes](https://help.getzep.com/memory-mcp-server "Zep: yes. Memory MCP Server, with sign-in through the company’s identity provider. (read on 2026-10-01)")

[partial](https://docs.letta.com/agent-sdk/mcp "Letta: partial. Letta connects to MCP servers as a client; it is not documented as a server. (read on 2026-10-01)")

[yes](https://supermemory.ai/docs/supermemory-mcp/mcp "Supermemory: yes. A hosted server; not available in the local version. (read on 2026-10-01)")

[yes](https://github.com/vectorize-io/hindsight "Hindsight: yes. A built-in MCP endpoint, one per bank. (read on 2026-10-01)")

[yes](https://honcho.dev/docs/v3/guides/integrations/mcp "Honcho: yes (read on 2026-10-01)")

[yes](https://docs.cognee.ai/cognee-mcp/mcp-overview "Cognee: yes (read on 2026-10-01)")

[partial](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-gateway-connector.html "Amazon Bedrock AgentCore Memory: partial. Through AgentCore Gateway, with a memory connector you set up. (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: not documented (read on 2026-10-01)")

A published benchmarkThe vendor publishes measured results of its own memory, with the method.

[yes](https://niadra.com/benchmark "Niadra: yes. Cross-channel accuracy, privacy, cost and latency, with a public script and result files. (read on 2026-10-01)")

[yes](https://mem0.ai/research "Mem0 (open source): yes. LoCoMo, LongMemEval and BEAM, with a paper on arXiv. (read on 2026-10-01)")

[yes](https://mem0.ai/research "Mem0 Platform: yes. LoCoMo, LongMemEval and BEAM, with a paper on arXiv. (read on 2026-10-01)")

[yes](https://arxiv.org/abs/2501.13956 "Zep: yes. DMR and LongMemEval, in Zep’s paper. (read on 2026-10-01)")

[yes](https://www.letta.com/blog/benchmarking-ai-agent-memory/ "Letta: yes. LoCoMo with Letta’s filesystem, on the blog. (read on 2026-10-01)")

[partial](https://supermemory.ai/docs/memorybench/overview "Supermemory: partial. Publishes the open MemoryBench harness; the results are yours to run. (read on 2026-10-01)")

[yes](https://arxiv.org/abs/2512.12818 "Hindsight: yes. LongMemEval, with a paper and live results. (read on 2026-10-01)")

[yes](https://honcho.dev/evals/ "Honcho: yes. LongMemEval, LoCoMo and BEAM. (read on 2026-10-01)")

[yes](https://github.com/topoteretes/cognee "Cognee: yes. BEAM, with a report in the repository and a paper on arXiv. (read on 2026-10-01)")

[not documented](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html "Amazon Bedrock AgentCore Memory: not documented. Says the strategies are \"benchmarked\"; no result published. (read on 2026-10-01)")

[not documented](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: not documented (read on 2026-10-01)")

Open source and self-hostingThe memory engine is open and can run on the customer’s own infrastructure.

[no](https://github.com/ainiadra "Niadra: no. The SDKs are open (Apache 2.0); the service is fully managed by Niadra, by the company’s decision. (read on 2026-10-01)")

[yes](https://github.com/mem0ai/mem0 "Mem0 (open source): yes. Apache 2.0, with a Docker server. (read on 2026-10-01)")

[no](https://mem0.ai/pricing "Mem0 Platform: no. A managed service; on-premises on Enterprise. (read on 2026-10-01)")

[partial](https://help.getzep.com/graphiti/getting-started/overview "Zep: partial. Graphiti is open; Zep Cloud is managed, with bring-your-own-cloud on Enterprise. (read on 2026-10-01)")

[yes](https://github.com/letta-ai/letta "Letta: yes. Apache 2.0, with its own server. (read on 2026-10-01)")

[yes](https://supermemory.ai/docs/self-hosting/overview "Supermemory: yes. An open local binary; no connectors and no MCP locally. (read on 2026-10-01)")

[yes](https://github.com/vectorize-io/hindsight "Hindsight: yes. MIT, one container. (read on 2026-10-01)")

[yes](https://github.com/plastic-labs/honcho "Honcho: yes. AGPL-3.0, with docker compose. (read on 2026-10-01)")

[yes](https://github.com/topoteretes/cognee "Cognee: yes. Apache 2.0, pip or Docker. (read on 2026-10-01)")

[no](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html "Amazon Bedrock AgentCore Memory: no. A managed AWS service. (read on 2026-10-01)")

[no](https://www.twilio.com/docs/conversations/memory "Twilio Conversation Memory: no. A managed Twilio service; not HIPAA eligible and not PCI compliant. (read on 2026-10-01)")

Published priceThe list price as the vendor’s page states it, with no conversion. The cost per thousand conversations, measured by the same method for every system, is on the benchmark.

US$ 2 to 3 per thousand conversations or tasks, models included; US$ 500 a month minimum[source · 2026-10-01](https://niadra.com/preco)

Free software; you pay for the model and the servers[source · 2026-10-01](https://github.com/mem0ai/mem0)

Hobby free; Starter US$ 19 a month; Pro US$ 249 a month (500,000 adds and 50,000 retrievals); Enterprise on request[source · 2026-10-01](https://mem0.ai/pricing)

Flex US$ 125 a month with 50,000 credits (1 credit per episode of up to 350 bytes); Flex Plus US$ 375; Enterprise on request[source · 2026-10-01](https://www.getzep.com/pricing)

API US$ 20 a month, plus US$ 0.10 per active agent a month and US$ 0.00015 per second of tool execution; the model apart[source · 2026-10-01](https://docs.letta.com/pricing)

Pro US$ 19, Max US$ 100 and Scale US$ 399 a month; rate card of US$ 5 per million SM tokens and US$ 5 per million searches[source · 2026-10-01](https://supermemory.ai/pricing)

Usage-based: retain US$ 10 per million tokens, recall US$ 0.75 per million, reflect US$ 0.05 per call, storage US$ 0.25 per million a month[source · 2026-10-01](https://vectorize.io/pricing)

Ingestion US$ 2 per million tokens; context() unlimited; reasoning from US$ 0.001 to 0.50 per question[source · 2026-10-01](https://honcho.dev/)

Free up to 1 million tokens; Standard US$ 1 per million tokens plus US$ 5 per extra workspace; Enterprise on request[source · 2026-10-01](https://www.cognee.ai/pricing)

US$ 0.25 per 1,000 events; US$ 0.75 per 1,000 records stored a month (US$ 0.25 with your own strategy); US$ 0.50 per 1,000 retrievals[source · 2026-10-01](https://aws.amazon.com/bedrock/agentcore/pricing/)

US$ 0.005 per profile used per day; US$ 0.0028 per 1,000 characters generated; US$ 0.007 per recall[source · 2026-10-01](https://www.twilio.com/en-us/products/conversational-ai/pricing)

Names and logos belong to their owners and appear here only to identify the systems compared.

[See the benchmark, with the measured numbers and the result files](/en/benchmark)

## The evidence behind every mark

System by system, what the page read says about each capability, with the link. It is what you need to check any cell.

### NiadraOpenClose

Identity across channels: yes

Identifiers by type, with a weight, reversible merges and a verification level per conversation.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/identity)

One memory for agents from different vendors: yes

Each vendor has its own credential and permission; none reads the other’s memory or numbers.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/integrations/overview)

Events from the CRM and the ERP: yes

Webhook with a versioned mapping, batch and file; no AI in the path.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/systems)

Typed business state with freshness: yes

Types declared or derived from the database; freshness per field computed on the read.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/object-types)

The claim check: yes

A contract per category; the check runs in the agent’s process, with no extra AI.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/claims)

Coordination between agents: yes

Ownership by declaration and observation, effect exactly once, a contact token for the gateway.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/coordination)

Turn records and replay: yes

Captured in the agent’s process; replay in CI with the tools answering from the record.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/turn-records)

Per-agent permissions and field access: yes

Denied by default, granted by purpose; fixed sensitive categories; field access on the state.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/privacy)

Receipts of who read what: yes

A receipt per read, chained by SHA-256, with a daily root in write-once storage.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/receipts)

Erasure with a receipt: yes

Erasure by lineage, with a receipt and the list of what to erase in your exported copy.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/privacy)

Prompt-ready compact context: yes

Compiled when the memory changes, per agent and task, in layers the provider reuses.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/concepts/context)

Voice integrations: yes

LiveKit, Pipecat, Vapi, Retell, ElevenLabs and Twilio, with the context before the hello.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/integrations/overview)

MCP server: yes

Eight tools, with the customer bound out of the model’s reach.

[source · 2026-10-01(opens docs.niadra.com)](https://docs.niadra.com/guides/mcp)

A published benchmark: yes

Cross-channel accuracy, privacy, cost and latency, with a public script and result files.

[source · 2026-10-01(opens niadra.com)](https://niadra.com/benchmark)

Open source and self-hosting: no

The SDKs are open (Apache 2.0); the service is fully managed by Niadra, by the company’s decision.

[source · 2026-10-01(opens github.com)](https://github.com/ainiadra)

### Mem0 (open source)OpenClose

Identity across channels: no

Memory is kept under the user\_id the application sends.

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

One memory for agents from different vendors: partial

Any agent with the key reads the same user\_id. No permission per vendor.

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

Events from the CRM and the ERP: not documented

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/open-source/overview)

Typed business state with freshness: no

Memory is text extracted from the conversation.

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

The claim check: not documented

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

Coordination between agents: not documented

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

Turn records and replay: not documented

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

Per-agent permissions and field access: partial

An API key per server user; no purpose and no field.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/open-source/features/rest-api)

Receipts of who read what: partial

A recent log of the server’s API calls.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/open-source/features/rest-api)

Erasure with a receipt: partial

Deletion per memory or user; no receipt.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/open-source/features/rest-api)

Prompt-ready compact context: partial

Search returns a list of memories; the application builds the prompt.

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

Voice integrations: yes

LiveKit, ElevenLabs and Pipecat.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/integrations/livekit)

MCP server: yes

OpenMemory MCP, local, on top of open-source Mem0.

[source · 2026-10-01(opens mem0.ai)](https://mem0.ai/blog/introducing-openmemory-mcp)

A published benchmark: yes

LoCoMo, LongMemEval and BEAM, with a paper on arXiv.

[source · 2026-10-01(opens mem0.ai)](https://mem0.ai/research)

Open source and self-hosting: yes

Apache 2.0, with a Docker server.

[source · 2026-10-01(opens github.com)](https://github.com/mem0ai/mem0)

### Mem0 PlatformOpenClose

Identity across channels: no

user\_id, agent\_id, app\_id and run\_id come from the application.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/features/entity-scoped-memory)

One memory for agents from different vendors: partial

Several agents filter the same user\_id by agent\_id. No permission per vendor.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/features/entity-scoped-memory)

Events from the CRM and the ERP: not documented

Mem0’s webhooks are outbound: they notify when a memory changes.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/features/webhooks)

Typed business state with freshness: partial

A creation date and an expiry per memory; no typed object and no freshness per field.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/features/timestamp)

The claim check: not documented

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/overview)

Coordination between agents: not documented

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/overview)

Turn records and replay: not documented

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/overview)

Per-agent permissions and field access: partial

Reader and owner roles per project; no purpose and no field.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/api-reference/organizations-projects)

Receipts of who read what: partial

A project event feed (add, search, delete); audit logs on Enterprise.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/platform-vs-oss)

Erasure with a receipt: partial

Deleting a user is asynchronous and returns an event\_id; no receipt.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/api-reference/entities/delete-user)

Prompt-ready compact context: partial

Search returns a list of memories; the application builds the prompt.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/overview)

Voice integrations: yes

LiveKit, ElevenLabs and Pipecat.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/integrations/livekit)

MCP server: yes

A hosted server with eleven tools.

[source · 2026-10-01(opens docs.mem0.ai)](https://docs.mem0.ai/platform/mem0-mcp)

A published benchmark: yes

LoCoMo, LongMemEval and BEAM, with a paper on arXiv.

[source · 2026-10-01(opens mem0.ai)](https://mem0.ai/research)

Open source and self-hosting: no

A managed service; on-premises on Enterprise.

[source · 2026-10-01(opens mem0.ai)](https://mem0.ai/pricing)

### ZepOpenClose

Identity across channels: no

The user\_id "maps to your internal User ID".

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/users-and-user-graphs)

One memory for agents from different vendors: partial

The company’s agent and the person’s MCP client share one user graph.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/memory-mcp-server)

Events from the CRM and the ERP: partial

Business records go in as JSON through graph.add; no inbound webhook and no per-system ID.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/adding-business-data)

Typed business state with freshness: partial

Facts with valid\_at and invalid\_at; no typed object per field.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/facts)

The claim check: not documented

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/concepts)

Coordination between agents: not documented

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/concepts)

Turn records and replay: not documented

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/concepts)

Per-agent permissions and field access: partial

Policies on API keys and user groups; no purpose and no field.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/governance)

Receipts of who read what: partial

Audit logs of API and dashboard activity; no receipt per read.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/governance)

Erasure with a receipt: partial

Deletes the user, the threads and the graph; returns a success message.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/sdk-reference/user/delete)

Prompt-ready compact context: yes

A context block assembled as text; ingestion can take minutes.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/retrieving-context)

Voice integrations: yes

ElevenLabs and LiveKit.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/memory-for-agent-frameworks)

MCP server: yes

Memory MCP Server, with sign-in through the company’s identity provider.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/memory-mcp-server)

A published benchmark: yes

DMR and LongMemEval, in Zep’s paper.

[source · 2026-10-01(opens arxiv.org)](https://arxiv.org/abs/2501.13956)

Open source and self-hosting: partial

Graphiti is open; Zep Cloud is managed, with bring-your-own-cloud on Enterprise.

[source · 2026-10-01(opens help.getzep.com)](https://help.getzep.com/graphiti/getting-started/overview)

### LettaOpenClose

Identity across channels: no

An identity carries the key the application chooses (identifier\_key).

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/api-reference/identities/create)

One memory for agents from different vendors: partial

Blocks shared between Letta agents, through git repositories. Other vendors are not mentioned.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/concepts/shared-memory)

Events from the CRM and the ERP: not documented

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/agent-sdk/memory)

Typed business state with freshness: no

Memory blocks in text, edited by the agent.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/agent-sdk/memory)

The claim check: not documented

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/agent-sdk/memory)

Coordination between agents: not documented

The shared memory page says nothing about locks or ownership.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/concepts/shared-memory)

Turn records and replay: partial

The steps API keeps the model, tokens and trace id of each step; replay is not documented.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/api-reference/steps/list)

Per-agent permissions and field access: partial

Role-based access control on the Enterprise plan.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/pricing)

Receipts of who read what: not documented

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/pricing)

Erasure with a receipt: partial

Deleting an agent returns an empty object.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/api-reference/agents/delete)

Prompt-ready compact context: partial

The blocks live in the Letta agent’s prompt; they are not served to another agent.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/agent-sdk/memory)

Voice integrations: yes

LiveKit and Vapi, marked experimental.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/guides/voice/livekit/)

MCP server: partial

Letta connects to MCP servers as a client; it is not documented as a server.

[source · 2026-10-01(opens docs.letta.com)](https://docs.letta.com/agent-sdk/mcp)

A published benchmark: yes

LoCoMo with Letta’s filesystem, on the blog.

[source · 2026-10-01(opens www.letta.com)](https://www.letta.com/blog/benchmarking-ai-agent-memory/)

Open source and self-hosting: yes

Apache 2.0, with its own server.

[source · 2026-10-01(opens github.com)](https://github.com/letta-ai/letta)

### SupermemoryOpenClose

Identity across channels: no

One profile per containerTag, set by the application.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/user-profiles)

One memory for agents from different vendors: partial

Keys restricted to containerTags; any agent with the key reads the profile.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/authentication)

Events from the CRM and the ERP: not documented

The connectors are knowledge sources: Drive, Gmail, Notion, OneDrive.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/connectors/google-drive)

Typed business state with freshness: partial

Facts with update relations and expiry; no typed object per field.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/graph-memory)

The claim check: not documented

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/graph-memory)

Coordination between agents: not documented

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/graph-memory)

Turn records and replay: not documented

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/graph-memory)

Per-agent permissions and field access: partial

Keys restricted to containerTags; no purpose and no field.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/authentication)

Receipts of who read what: not documented

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/authentication)

Erasure with a receipt: partial

Forgetting is a soft delete; it returns the id and the forgotten flag.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/api-reference/content-management/forget-a-memory)

Prompt-ready compact context: yes

A static and a dynamic profile, for the prompt.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/concepts/user-profiles)

Voice integrations: yes

LiveKit and Pipecat.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/integrations/livekit)

MCP server: yes

A hosted server; not available in the local version.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/supermemory-mcp/mcp)

A published benchmark: partial

Publishes the open MemoryBench harness; the results are yours to run.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/memorybench/overview)

Open source and self-hosting: yes

An open local binary; no connectors and no MCP locally.

[source · 2026-10-01(opens supermemory.ai)](https://supermemory.ai/docs/self-hosting/overview)

### HindsightOpenClose

Identity across channels: no

One memory bank per user, project or agent, isolated from the others.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/memory-banks)

One memory for agents from different vendors: partial

Keys restricted to banks; any agent with the key reads the bank.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/api-keys)

Events from the CRM and the ERP: not documented

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/retain)

Typed business state with freshness: no

Text items, with an optional timestamp.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/retain)

The claim check: not documented

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/retain)

Coordination between agents: not documented

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/memory-banks)

Turn records and replay: not documented

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/retain)

Per-agent permissions and field access: partial

Keys restricted to banks; no purpose and no field.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/api-keys)

Receipts of who read what: not documented

The SIEM receives only Memory Defense violations, with the key that wrote.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/security/webhooks-siem)

Erasure with a receipt: yes

Two-step erasure with a receipt in the privacy events ledger; Enterprise only.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/security/data-privacy)

Prompt-ready compact context: partial

recall returns memories; reflect returns an answer a model writes on the read.

[source · 2026-10-01(opens docs.hindsight.vectorize.io)](https://docs.hindsight.vectorize.io/recall)

Voice integrations: yes

Pipecat, Vapi and ElevenLabs.

[source · 2026-10-01(opens github.com)](https://github.com/vectorize-io/hindsight)

MCP server: yes

A built-in MCP endpoint, one per bank.

[source · 2026-10-01(opens github.com)](https://github.com/vectorize-io/hindsight)

A published benchmark: yes

LongMemEval, with a paper and live results.

[source · 2026-10-01(opens arxiv.org)](https://arxiv.org/abs/2512.12818)

Open source and self-hosting: yes

MIT, one container.

[source · 2026-10-01(opens github.com)](https://github.com/vectorize-io/hindsight)

### HonchoOpenClose

Identity across channels: no

Each peer carries the identifier the application gives it.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

One memory for agents from different vendors: partial

Several peers in a session, with a setting for who observes whom.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Events from the CRM and the ERP: not documented

Honcho’s webhooks are outbound: they notify when the queue empties.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/features/advanced/webhooks)

Typed business state with freshness: no

Representations and conclusions about the person, in text.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

The claim check: not documented

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Coordination between agents: not documented

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Turn records and replay: not documented

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Per-agent permissions and field access: partial

A setting for which peers observe which; no purpose and no field.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Receipts of who read what: not documented

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/features/advanced/deleting-data)

Erasure with a receipt: partial

Asynchronous deletion per session; derived conclusions survive; no receipt.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/features/advanced/deleting-data)

Prompt-ready compact context: yes

get\_context assembles conclusions and recent messages for the prompt.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/documentation/core-concepts/architecture)

Voice integrations: not documented

The documented integrations are LangGraph, CrewAI, Vercel AI SDK and n8n.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/llms.txt)

MCP server: yes

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/docs/v3/guides/integrations/mcp)

A published benchmark: yes

LongMemEval, LoCoMo and BEAM.

[source · 2026-10-01(opens honcho.dev)](https://honcho.dev/evals/)

Open source and self-hosting: yes

AGPL-3.0, with docker compose.

[source · 2026-10-01(opens github.com)](https://github.com/plastic-labs/honcho)

### CogneeOpenClose

Identity across channels: no

Users and tenants come from the application; memory is per dataset.

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview)

One memory for agents from different vendors: partial

Datasets shared through an access list (read, write, delete, share).

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview)

Events from the CRM and the ERP: not documented

The data sources are Slack, Notion, Linear and Google Drive.

[source · 2026-10-01(opens www.cognee.ai)](https://www.cognee.ai/pricing)

Typed business state with freshness: partial

Events with timestamps and fact validity; no business object per field.

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/guides/time-awareness)

The claim check: not documented

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

Coordination between agents: not documented

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

Turn records and replay: not documented

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

Per-agent permissions and field access: partial

Read, write, delete and share per dataset; no agent and no field.

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview)

Receipts of who read what: not documented

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview)

Erasure with a receipt: partial

forget returns ids and a status; no receipt.

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/core-concepts/main-operations/forget)

Prompt-ready compact context: partial

Search answers with a model at read time.

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

Voice integrations: not documented

The documented integrations are coding assistants and MCP.

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

MCP server: yes

[source · 2026-10-01(opens docs.cognee.ai)](https://docs.cognee.ai/cognee-mcp/mcp-overview)

A published benchmark: yes

BEAM, with a report in the repository and a paper on arXiv.

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

Open source and self-hosting: yes

Apache 2.0, pip or Docker.

[source · 2026-10-01(opens github.com)](https://github.com/topoteretes/cognee)

### Amazon Bedrock AgentCore MemoryOpenClose

Identity across channels: no

The actorId is set by the application on every event.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html)

One memory for agents from different vendors: partial

Any IAM principal, even from another AWS account, reads the same memory.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-cross-account-access.html)

Events from the CRM and the ERP: not documented

An event is a message, a system action or a tool call of the conversation.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-terminology.html)

Typed business state with freshness: no

Facts, preferences and summaries extracted from the conversation.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html)

The claim check: not documented

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html)

Coordination between agents: not documented

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html)

Turn records and replay: partial

Extraction metrics and logs in CloudWatch; replay is not documented.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-observability.html)

Per-agent permissions and field access: partial

IAM policies by namespace and actor; no field access.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html)

Receipts of who read what: partial

CloudTrail logs the API calls; no chained receipt per read.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-cross-account-access.html)

Erasure with a receipt: partial

Deleting a record returns its id; no receipt.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/APIReference/API_DeleteMemoryRecord.html)

Prompt-ready compact context: partial

Retrieval returns records; the application builds the prompt.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-organization.html)

Voice integrations: not documented

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html)

MCP server: partial

Through AgentCore Gateway, with a memory connector you set up.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-gateway-connector.html)

A published benchmark: not documented

Says the strategies are "benchmarked"; no result published.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html)

Open source and self-hosting: no

A managed AWS service.

[source · 2026-10-01(opens docs.aws.amazon.com)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html)

### ![](/integrations/logos/twilio.png)Twilio Conversation MemoryOpenClose

Identity across channels: yes

Phone, e-mail, WhatsApp and user ID become one profile, only for conversations that flow through Twilio’s channels.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/identity-resolution)

One memory for agents from different vendors: partial

Any agent calls the recall API with the account key; the conversation must flow through Twilio.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/api-quickstart)

Events from the CRM and the ERP: partial

Profiles and traits go in through bulk import; no order or bill event.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/api-quickstart)

Typed business state with freshness: partial

Factual memory as key-value traits; no freshness per field.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

The claim check: not documented

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

Coordination between agents: not documented

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

Turn records and replay: not documented

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

Per-agent permissions and field access: not documented

Authentication with the account key; no permission per agent.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/api-quickstart)

Receipts of who read what: partial

Error and warning logs of the operations, retrievals included.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/observability-logs)

Erasure with a receipt: partial

Deleting a profile is accepted (202) and irreversible; no receipt.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/api/memory/v1/profile/delete-profile)

Prompt-ready compact context: partial

Recall returns structured observations and summaries; the application builds the prompt.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory/api-quickstart)

Voice integrations: yes

Voice through Twilio itself, with ConversationRelay.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/en-us/blog/developers/tutorials/product/orchestrate-multi-call-conversations-with-llm-twilio-conversation-memory)

MCP server: not documented

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

A published benchmark: not documented

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

Open source and self-hosting: no

A managed Twilio service; not HIPAA eligible and not PCI compliant.

[source · 2026-10-01(opens www.twilio.com)](https://www.twilio.com/docs/conversations/memory)

What the others have and Niadra does not: open source and self-hosting. Niadra is fully managed by the company’s decision, and the SDKs are open. Found a wrong cell or a newer page? Use the form on the Enterprise page: the cell changes, with the date of the reading.

[See the benchmark, with the measured numbers and the result files](/en/benchmark)

## Tell us what you are building.

A work email and two lines about your agents are enough. The people who write the code reply, with an early-access proposal for your case.

[Rather tell us more about your company? Use the full form](/en/enterprise)