Agent Memory Systems

Memory can include recent conversation state, selected facts, preferences, past events or summaries of prior work. A memory system has write, update, retrieval and deletion behavior, often using structured stores, embeddings or graph relationships. Some systems extract candidate facts from interactions before storing them. This differs from simply appending all messages to a context window and from checkpointing a run for recovery. Memory changes the information available to future decisions; it does not necessarily update model weights. The design must represent provenance, age and ownership because useful recollection and accidental persistence of a mistaken statement can look similar to a generator.

The practitioner defines retention criteria, user boundaries and how conflicting or obsolete records are handled. Retrieval should select relevant memory rather than injecting everything available. Explicit user corrections and deletion requests need reliable update paths. Useful artifacts include a memory schema, provenance records and tests for cross-user isolation and stale facts. Evaluation asks whether memory improves continuity on future tasks and whether incorrect retention causes failures. Storage and disclosure rules should match the purpose for which the information was collected.

Extracted memories can flatten qualifications, preserve temporary preferences or amplify an early error. Summaries may lose the evidence needed to resolve contradictions. More memory is not automatically more useful, and personalization can become intrusive if retention is not visible or controllable. Quality requires relevant retrieval, accurate updates, provenance and deletion behavior. Long-term memory should not be confused with current external truth: an old remembered schedule still needs verification before it supports a new decision.