CORE GUIDE
Memory lifecycle
Durable AI memory needs admission, provenance, update, retrieval and deletion rules; persistence without a lifecycle only makes stale mistakes harder to remove.
Mental model
Memory is managed state, not an append-only transcript. A durable memory item should have a reason to exist, a source, an owner or scope, a validity horizon and a policy for replacement or deletion.
Why it matters
Long-lived assistants become unreliable when every observation is remembered forever. Temporary preferences become permanent, incorrect model inferences harden into facts and old workflow state competes with current evidence. A lifecycle keeps memory useful by distinguishing what deserves persistence from what should remain working context or be re-derived from an authoritative source.
01
Manage memory from write decision to retirement
Before writing, classify the candidate memory: verified fact, user preference, workflow checkpoint or derived summary. Store provenance, scope and timestamps with the value. At retrieval time, select memories relevant to the current task and check authority or freshness. Update or supersede values when stronger evidence arrives, and expire or delete memories when their purpose, consent or validity ends.
02
Example: a user's preferred report format changes
The assistant stores a verified preference for concise weekly reports with the user's scope and update time. Months later the user requests detailed reports. The new explicit preference supersedes the old record instead of adding a contradictory second memory. Future tasks retrieve only the current preference while the audit history remains separate if needed.
Common failure modes
- Persisting raw model guesses as durable facts.
- Appending contradictory memories without supersession or authority rules.
- Retrieving all historical memories into every prompt regardless of task relevance.
Engineering heuristics
- Require a memory type and provenance before durable writes.
- Separate current effective state from historical audit history.
- Give every durable memory class an update, expiry or deletion policy.
Takeaways
- 01Persistence is useful only with lifecycle control.
- 02Memory should be selected and invalidated, not merely accumulated.
- 03Authoritative current state must outrank old conversational residue.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.