1. – Independent Researcher, Redmond, WA, USA.
| Received
24-Sep-2026 |
Accepted
03-Oct-2026 |
Published
15-Oct-2026 |
Abstract
Persistent memory lets an agent act across sessions, but persistence alone does not make remembered state correct. Contemporary systems increasingly transform raw interactions into episodic narratives, reflections, summaries, graph facts, long-term profiles or tiered memory objects. These transformations create derived state. A correction to an earlier observation can therefore leave a semantically stale summary or inferred fact reachable even when the source record has changed. Existing surveys organise memory by representation, function, lifecycle stage or multi-agent sharing, while current benchmarks mainly measure recall, reasoning, capacity and efficiency. This structured narrative review studies a narrower systems question: how should an agent preserve semantic coherence when memory is consolidated, updated,
forgotten, moved across tiers or reconstructed after failure? We synthesise agent-memory work with concepts from replicated storage, materialised-view maintenance and human memory consolidation. The analysis separates retrieval relevance from state correctness and proposes a lineage-aware coherence contract in which derived memories carry identity, dependencies, version or validity information and explicit invalidation behaviour. The contract is an author-derived design, not an evaluated implementation. The review finds strong mechanisms for memory formation and retrieval, growing support for update and deletion operations and comparatively limited evaluation of stale derivatives, invalidation propagation, repair or selective forgetting. These gaps motivate benchmarks that inject corrections and revocations into long-running memory histories and measure whether every dependent representation converges on the new meaning.
Locked
Subscribed
Open Access
Locked Content