Long-Term Memory

Persistent state that an agent can retrieve across sessions, tasks, or executions. It may store preferences, confirmed facts, decisions, summarized episodes, or procedures, usually outside the immediate context in a searchable store. Persistence belongs to the architecture around the model and does not imply weight updates.

The central challenge is deciding what deserves to survive. Writes need origin, consent, scope, expiration, and a correction mechanism; reads need relevance and access control. Memory without policy accumulates stale or sensitive information and returns plausible context that should no longer govern a decision.