Lab

Capturing explanations without overloading teams

Short answer

Useful memory begins with an explanation someone is willing to give. Asking too many questions can undermine the work the feature was meant to support. The Context integration studied with Tsuno therefore treats human attention as a limited resource, alongside computation.

Selected events lead to a few explanation requests and reusable memory.

Ask for the reason when it matters

Not every business event deserves a new note. An isolated anomaly may have little future value. A recurring problem or a decision that commits a team may justify preserving an explanation.

The starting point is straightforward: operational facts are not enough to reconstruct intent. A price increase does not explain why a supplier was retained. An incident does not explain why repair was preferred to replacement. These examples illustrate a type of context; they do not describe actual customers.

A useful question asks for what might otherwise be missing later. It should not require someone to repeat information the application already holds. That distinction keeps memory collection connected to an actual knowledge gap.

Silence is part of the design

The policy explored in Tsuno combines several conditions: the event’s importance, recurrence, the absence of current context and the person’s permission to respond. When a relevant explanation already exists, the first step is to read or review it.

Requests are also limited and grouped. A person can answer, postpone, dismiss a question or say that no explanation exists. They can indicate that the question belongs with someone else. None of these choices should block an operational task.

Restraint affects information quality. Repetitive questions can lead to rushed answers and weaker memory. Collecting more text does not necessarily create better future context.

Relevance should therefore be assessed partly through silence: which unnecessary questions were avoided, and which questions actually preserved a reason that could be reused? A high response count alone would be an incomplete success measure.

Keep the human explanation recognisable

Before recording, a person should understand the exact text being sent, the business object involved and who can see it. Assisted rewriting should preserve attribution and avoid turning a hypothesis into a decision.

The contract distinguishes the type of statement, its certainty and any validity or review dates. Those qualifications help later readers interpret an explanation without granting it more authority than it originally had.

A correction or replacement decision should retain its relationship with the earlier version. Simply accumulating versions would leave the assistant to choose between potentially incompatible positions without knowing how they relate.

Different interfaces, shared responsibilities

The HTTPS API and SDK let an application record and retrieve context through its own screens and workflows. MCP provides an interface for authorised assistants. An event stream helps the publisher follow the lifecycle changes covered by the contract.

None of these interfaces changes the central rule: retrieved memory is data. Its content does not become an instruction to execute. The responsible applications must calculate and check permissions even when a passage appears relevant.

The application should also describe the retrieval honestly: memory was read, no memory was found, or memory was unavailable. An empty result and a service failure mean different things to someone making a decision.

Measure use beyond the form: scope and limits

Integration tests check the path from an explanation to its retrieval, including permission and availability cases. They do not establish that people will willingly provide explanations during everyday work.

That requires observing effort, dismissals and reuse, without collecting unnecessary content merely to measure activity. The synthetic evaluation results explain what useful context can contribute. Object-linked business memory explains where to retrieve it. Restraint helps determine whether someone will want to create it in the first place.

Related reading: What should AI remember from a meeting, project or decision? and Why a good memory must also know when not to write.

Your next prompt should not have to contain your whole story.

Start building a memory your AI tools can find again when it matters.

Start free