{"schemaVersion":"1.0.0","id":"a-better-home-for-agent-memory","slug":"a-better-home-for-agent-memory","title":"A better home for your agents’ memory","description":"Separate facts, working context, and durable decisions before adding another vector database.","category":"Field notes","tags":["Agents","Memory","Architecture"],"date":"2026-09-29","updated":"2026-09-29","author":"PlainNerd editorial","readTime":3,"illustrationKind":"workflow","status":"published","evidenceStatus":"research-plan","locale":"en","originalLocale":"en","content":{"type":"doc","content":[{"type":"paragraph","content":[{"type":"text","text":"Research plan. These experiments have not been run; no measured results are claimed."}]},{"type":"heading","attrs":{"level":2},"content":[{"type":"text","text":"Memory needs a job description"}]},{"type":"paragraph","content":[{"type":"text","text":"An agent can need several different kinds of information: temporary task context, durable user preferences, source documents, and records of decisions. Putting everything into one undifferentiated store makes retrieval and correction difficult to reason about."}]},{"type":"heading","attrs":{"level":2},"content":[{"type":"text","text":"Design a small memory record"}]},{"type":"paragraph","content":[{"type":"text","text":"Our proposed record includes the claim, its source, when it was observed, who may use it, and when it should be reviewed. Treat inferred preferences as tentative. A correction should supersede an earlier record rather than silently leaving both active."}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","content":[{"type":"text","text":"Facts need provenance and an explicit owner."}]}]},{"type":"listItem","content":[{"type":"paragraph","content":[{"type":"text","text":"Decisions need the context in which they were made."}]}]},{"type":"listItem","content":[{"type":"paragraph","content":[{"type":"text","text":"Working notes need an expiry or a clear task boundary."}]}]},{"type":"listItem","content":[{"type":"paragraph","content":[{"type":"text","text":"Sensitive material needs an intentional retention policy."}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"type":"text","text":"Do not confuse caching with remembering"}]},{"type":"paragraph","content":[{"type":"text","text":"Prompt caching reuses previously processed prompt prefixes when the provider’s cache requirements are met. It is a different mechanism from deciding which facts should persist across tasks. Cache hits can reduce repeated processing costs, while cache-write costs, lifetime, and model support affect the overall saving. Caching does not make the context accurate or appropriate."}]},{"type":"paragraph","content":[{"type":"text","text":"Anthropic: prompt caching","marks":[{"type":"link","attrs":{"href":"https://platform.claude.com/docs/en/build-with-claude/prompt-caching"}}]}]},{"type":"heading","attrs":{"level":2},"content":[{"type":"text","text":"Test correction, not just recall"}]},{"type":"paragraph","content":[{"type":"text","text":"A useful future experiment would ask an agent to retrieve a fact, incorporate a correction, and stop relying on the earlier value. Include cases where the right answer is that the source is missing or out of date. We have not run this evaluation; the record format and test cases are a research proposal."}]}]}}