A memory module gives a cognitive agent persistent long-term memory. It retrieves relevant prior experience before planning, records outcomes after evaluation, and turns separate tasks into cumulative institutional knowledge. section 10.2.8 section 10.2.1
Treat the module as an active layer over a structured knowledge graph:
- Proactive retrieval searches for entities related to the current task and injects matching observations into the workspace before planning. section 10.2.8
- Experience recording stores the completed task, strategy, and evaluation outcome, adding observations to an existing problem type or creating a new entity. section 10.2.8
- Relation building connects related problem types, successful tool combinations, and known failure modes when strategies succeed across different tasks. section 10.2.8
- The underlying graph uses entities as primary nodes, relations as directed connections, and observations as discrete information attached to entities. section 10.2.8
Retrieving experience during diagnosis
For an intermittent deployment failure occurring only under high traffic, memory retrieval can find a prior connection-pool exhaustion entity with an observation about failures under load and increasing max_connections. Planning then tests connection exhaustion, health-check races, and resource contention instead of repeating a generic timeout fix. section 10.3.3
