Memory is a product database with retention rules, not a large chat transcript you keep appending to forever.
Why it matters
Useful agents need continuity: user preferences, workspace facts, decisions, and prior outcomes. But memory that is broad, stale, or unsourced makes the agent confidently wrong. The design problem is deciding what deserves persistence and how it earns attention later.
Build this
- A short-term state store for the active task, separate from long-term facts.
- Memory records with owner, scope, source, created date, last validated date, and confidence.
- Retrieval rules that load only memories relevant to the current task and user.
- Correction, expiry, and deletion flows when memory is wrong or no longer allowed.
Watch for
- Storing every conversation turn as long-term truth.
- Loading stale preferences after the user has changed their mind.
- Sharing private agent memory across users, tenants, or scopes.
- Letting generated summaries replace primary source records.
Proof it works
- The agent can explain which memory influenced an answer.
- A changed preference overrides the older memory in the next relevant task.
- Private, tenant, and shared memories have separate access tests.
Implementation checklist
01
Classify memories as session, project, user, organisation, or global before storing them.
02
Attach memory to primary sources where possible rather than storing unsupported claims.
03
Cap memory retrieval by task relevance and attention budget.
04
Build a review surface for editing or retiring important memories.