Memory and retrieval for internal knowledge

Put approved policies, manuals, project notes, and runbooks in a document corpus with stable source identifiers. Use entity-scoped memory for information tied to a person or workflow, such as a preferred summary format or a reviewed decision from earlier work. Keeping those layers distinct makes ownership clearer and prevents a personal observation from being presented as organizational policy.

When both layers are relevant, retrieve them independently and label them in the prompt. The answer can then distinguish a source passage from remembered context. If they conflict, the current authoritative document should win or the workflow should ask for review instead of quietly selecting whichever text ranked higher.

Choose a small source set with known owners and update schedules. Preserve titles, source URLs or paths, modification dates, and any identifiers needed to replace or delete a document later. Chunking and embeddings make retrieval possible, but they do not tell the application whether the source is approved, expired, duplicated, or superseded.

Monitor accepted, failed, and pending ingestion separately from search quality. A successful upload does not prove that every useful passage is retrievable. Run representative questions after each material corpus change, and keep a path back to the source so a person can confirm the surrounding context before relying on an answer.

A search result should never broaden what the authenticated user may read. Map the caller’s workspace and role to the permitted corpus or partition on a trusted server, then issue the scoped query. Do not ingest a restricted document into a broadly shared index and rely on prompt instructions to hide it after retrieval.

When source permissions are complex or change frequently, consider separate partitions or indexes that follow the same boundary as the source system. Record the source identity and permission context used during ingestion so an access change can be reconciled. Aether provides retrieval primitives; mirroring an external content system’s authorization model remains integration work.

Pass only a bounded set of relevant passages to the model and keep source labels attached. Ask the agent to say when the available material does not answer the question. A link to the underlying policy or runbook lets the user inspect context, while a clear timestamp helps them recognize material that may need confirmation.

Memory can personalize the presentation without rewriting the evidence. For example, it may recall that one team prefers a short checklist, while the checklist items still come from the current runbook. This division preserves a useful experience and keeps organizational claims traceable to maintained documents.

Start with a bounded collection whose owners can judge the answers. Track retrieval relevance, unsupported answers, missing sources, permission failures, and stale documents. Add a route for users to report a bad source or request correction. Expand only after the ingestion, authorization, and review paths work together under realistic changes.