AI agent memory: writing, updating, forgetting and verification

Agent memory should make information reusable without losing its origin, scope or date. Storing every message in a vector database does not define how a fact will be corrected, shared or withdrawn. The design needs both a write path and a read path.

MemGPT studies memory tiers beyond directly available context. GAM separates historical preservation from on-demand research. These architectures inform choices while leaving application-specific permissions and lifecycle rules to define.

Separate information with different lifetimes

Layer	Fictional example	Reuse rule
Task state	A check is blocked by a missing account	Resume the current task state
Document fact	Procedure RZ-41 has revision 2	Check source and validity
Preference	A recipient wants a short summary	Respect scope and permit changes
Observation	An HTTP response was obtained at a given time	Preserve artefact and preconditions

Mixing these layers can turn a presentation preference into business policy or a temporary observation into permanent system state. Explicit types help determine what to verify before reuse.

Proposed writes need a status

Our proposed workflow stores candidate memories before promoting shared facts. Check provenance, identity, permissions, duplicates and contradictions. Some observations can be recorded immediately with their status; a consequential shared rule needs suitable promotion controls.

{
  "key": "procedure.vpn.window",
  "organisation": "alpha",
  "value": "Tuesday",
  "source": "procedure:RZ-41@2",
  "status": "confirmed",
  "supersedes": "procedure:RZ-41@1"
}

This educational record illustrates information missing from the plain statement “VPN maintenance is on Tuesday.” It is not a production implementation.

Resolve revisions without hiding conflicts

An old procedure says Friday; the new revision says Tuesday. Semantic similarity cannot choose between them. Apply revision and validity rules. Conflicting current sources should remain visible or trigger a decision request.

The last sentence in a conversation is not necessarily authoritative: it may be a question, historical quotation or unapproved proposal. Retain the event and evidence that changed a record’s status.

Expiry can prohibit use as current information while preserving authorised history. Deletion has a different purpose. Define the effects on embeddings, summaries, caches and secondary indexes in either case.

Test by recording a fact, producing a summary and revoking a user’s access. Check that a fresh consultation cannot recover it through the summary or cache. Repeat for deletion and direct identifier access using demonstration data.

Evaluate memory operations

MemSkill treats memory operations as skills whose selection and content can evolve. An operational question follows: who may change a write rule, and how can its effects be observed?

LongMemEval distinguishes abilities including temporal reasoning, updating and abstention. Adapt these dimensions to your service. Retrieving an old fact successfully does not establish recognition that it is now outdated.

The offline example provides deterministic revision, expiry and access checks, not an agent-memory implementation. Continue with context compaction and RAG security through the series index.