Research plan. These experiments have not been run; no measured results are claimed.
Memory needs a job description
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.
Design a small memory record
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.
Facts need provenance and an explicit owner.
Decisions need the context in which they were made.
Working notes need an expiry or a clear task boundary.
Sensitive material needs an intentional retention policy.
Do not confuse caching with remembering
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.
Test correction, not just recall
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 Include cases where the right answer is that the source is missing or out of date.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.
Prepared with AI assistance. Sources are linked in the story. This is a research plan, not a report of completed experiments. No measured results are claimed.
The conversation
MODERATED, ALWAYSA useful question, a different result, a missing detail. Start there.