30 June 2025 / Applied AI / 8 chapters

Decide which kinds of memory the system keeps

From Long-term AI memory: records, relationships and retrieval

Working context, durable facts, decisions, preferences and archives have different lifetimes and retrieval rules. Define those categories before choosing a database or embedding model, then decide which can share storage without sharing retention or retrieval behaviour.

Working context belongs to the active task. It includes the immediate objective, open plan, recent tool results and temporary assumptions. It should be compacted or cleared when the task ends. Promoting all of it into durable memory would preserve abandoned approaches and routine execution noise.

Durable factual records describe entities and their current properties: a person's approved contact preference, a project's status, a system's repository location or a service's owner. Each needs a source and a correction path. Some are time-bounded. A current owner may be valid from one date until another assignment replaces it.

Keep decisions in their own record type with the choice, scope, date, participants or authority, rationale where available and the alternatives affected. A decision may later be reversed while remaining historically important.

Preferences need careful scope. "Use short updates for this project" is different from a permanent personal preference. Store who expressed it, which context it applies to and whether it was explicit or inferred. Inferred preferences should have lower authority and be easy to remove.

Events record that something happened: a deployment, approval, handoff or failed run. Keep them as chronology and audit records. A controlled write can use a deployment event to update a separate current production version record.

The archive holds source documents, transcripts and older records for search and verification. It can be large because the default retrieval path does not pour it into every session.

All articles