The Boundaries of Trust

Hindsight can preserve source text and still return a missing, stale, or misleading answer. The important diagnostic question is not simply whether the bank contains the document, but which transition failed: extraction, relationship building, retrieval, consolidation, scope enforcement, or security screening. This chapter separates those symptoms, identifies the evidence that can distinguish them, and matches each cause with a recovery that changes derived knowledge without unnecessarily destroying the source record.

A bad answer is a symptom, not a diagnosis

Consider a user asking about the retry-policy incident described earlier. The agent answers that it has no information. Several different situations could produce that sentence.

The input may never have become a memory unit. A narrow retain mission can exclude every fact in a document while the document body remains stored. From the document list, the source appears to exist; from recall and reflect, it is unreachable because those operations search memory units. The right response is to inspect extraction outcome and reprocess under a broader mission, not to tune the reranker.

The fact may exist but carry the wrong interpretation. The speaker may have been classified as the agent rather than the engineer. A relative date may have been resolved against the wrong event reference. A screenshot may contain the request identifier, but the fact may not be attributed to that attachment. An entity may have merged with a similar name or failed to connect to the project. In those cases, the source is present, but the representation that retrieval sees is distorted.

The facts may be correct while the query misses them. Approximate vector search can omit a candidate. Keyword search may not recognize a spelling variation. A graph path may be absent because an entity link was not created. A time window may be too narrow. A low search budget or token limit may leave out a qualification. Increasing the number of model calls at the reflect layer will not fix an evidence set that recall never found.

The raw facts may be available while the observation is stale. Consolidation is asynchronous; a new incident can arrive after the last observation refresh. A mental model can preserve an older answer because its refresh has not run, because its source tags exclude the new fact, or because its delta baseline has drifted. In this case, asking for raw evidence may reveal the new event while the higher-level projection still says nothing about it.

Finally, the memory tools may fail as infrastructure. A database, embedder, or reranker error does not mean the bank lacks knowledge. Reflect treats a raised tool failure differently from an ordinary empty result precisely because the evidence path is incomplete. The diagnostic response should preserve the distinction between “no matching memory,” “memory not yet derived,” and “retrieval service failed.”