Measurement
You see whether each agent used the context it received.
Niadra delivered the context and receives the conversation that came next. That is why it can measure context use, per agent and per vendor: what the agent used, what it asked again, what it said against what the memory already knew and whether it transferred the conversation without the destination reading. No agent vendor can give you that number, because none of them holds both sides.
How it works
Context use, per agent
example| Agent | Used the context | Repeated a question | Contradictions |
|---|---|---|---|
| Voice agentvendor B | 91% | 3% | 0 |
| WhatsApp agentvendor A | 74% | 18% | 2 |
| Billing agentinternal | 88% | 1% | 0 |
Next to each agent, the 72-hour recontact rate. Every number opens the conversations and the receipt of each delivery.
The four signals
- Use
- The agent cited or acted on what it received: the missed visit, the credit applied, the earlier promise.
- Does not count: an item that did not serve the task. It counts as unused, with no penalty.
- Repeat
- The agent asked for data that was already in the context: name, ID number, order number, the reason for contact.
- Does not count: a verification question your policy requires.
- Contradiction
- The agent said something a delivered item disproves, such as “the credit is still under review” after the credit was applied.
- Does not count: a divergence from what was not delivered. That is a selection failure, measured separately.
- Transfer without reading
- The conversation was transferred to another agent or to a person, and the destination did not request the context.
- Does not count: a destination that is not integrated yet. That shows up as missing coverage.
- No extra AI model: it is the delivery receipt crossed with what the memory already extracts from the conversation
- Ready in under a minute after the conversation ends
- Vendor A never sees vendor B’s numbers
- It measures memory use, not service quality: it rates no one and blocks nothing
- The 90-day goal is recontact, against a control group. The measurement shows where it starts, and it goes into a contract only once sample review shows 90% agreement or more
One memory, nine products
Connect
What goes into the memory
SDK, API and MCP
Any agent connects to the memory with three calls, in Python, TypeScript, HTTP or MCP.
See the documentationRemember
What the memory knows
Deliver
What each agent receives
Govern
What your team controls
- MeasurementYou are hereWhether each agent used the context it received.
- ConsoleWho read what, review, export and erasure.