Glossary. The terms, in one sentence.

The record of every read: which agent read what, why and when. At Niadra it is chained with SHA-256 hashes and streamed to the company's SIEM.

When the customer switches channels in the middle of an issue, for example from a call to WhatsApp. With shared memory, the next agent starts out knowing what was handled.

How much the conversation has already proven who the person is. A phone number is a hint; a password login is proof. The context delivered changes with the level.

What the AI agent receives before it replies: who the customer is, what is still open, what just happened on another channel and what the history teaches, ready for the prompt.

Roles under the GDPR and the LGPD. The company that serves the customer is the controller; Niadra processes data on its behalf, as processor, under a data processing agreement (DPA).

A hardware module that keeps encryption keys without ever letting them out. FIPS 140-3 is the NIST standard for these modules; Niadra uses level 3.

Everything that has happened with the customer, open for the agent to search during the conversation when the context is not enough.

The lookup the agent runs during the conversation, by meaning and by keyword, filtered by period, channel and topic. Each result comes back with date, channel and outcome.

Passing the conversation from an AI agent to a human agent. With shared memory, the human gets the full case and does not ask for the tax ID again.

Working out that a phone number, an email, a tax ID and an app login belong to the same person, without relying on a user ID handed over ready-made by each channel.

Ordering the context from what changes least to what changes most (company rules, profile, open items, just now), so the provider can reuse the beginning.

Model Context Protocol, the open protocol that connects tools to AI models. Niadra's history search works as an MCP tool, in addition to the SDK and the HTTP API.

Everything Niadra keeps and organizes about a customer: full conversations, facts, preferences, open items and promises, with date and source.

A memory that is not an agent, not a channel and does not compete with agent vendors. That is why all of them can read and feed the same memory, and switching vendors erases nothing.

The customer memory that all of a company's AI agents share, on any channel and from any vendor. It keeps conversations, facts, open items and promises, with the source of each one.

The AI provider reusing the beginning of a prompt that repeats. The reused part costs much less, as long as it passes each model's minimum length.

The rule that sets what each agent may read based on the reason for the interaction. The collections agent never sees a technical issue or health data.

A phone number that changed owners. Treated as a hint, not as proof, it never hands one person's history to someone else.