Context resets
A new chat, model run, or coding session often starts without the source material that made the last answer useful.
Agent memory is the layer that lets an AI system reuse grounded facts across sessions, tools, and clients. It is not just a longer prompt or a vector search box; useful memory records provenance, supports deletion, handles changing facts, and lets humans inspect why an answer was recalled.
Forgetting is usually not a model defect. It is a product boundary: the agent lacks a durable, scoped, inspectable layer for facts that should survive one session.
A new chat, model run, or coding session often starts without the source material that made the last answer useful.
Docs, tickets, and decisions change. A memory layer needs supersession signals so stale facts do not look equally current.
Raw retrieval can return snippets without enough lineage for a human to inspect what the agent relied on.
The terms overlap in conversation, but buyers evaluate different primitives. Memory is the persistence and governance layer around recall.
Temporary working memory inside one model call or chat. Useful, but it disappears or grows expensive quickly.
Search over source chunks. Useful for recall, but not automatically a durable, inspectable memory system.
A persistent layer that stores scoped evidence, provenance, updates, and deletion semantics for agents to reuse.
Memory has to be inspectable because the user needs to know what was stored, why it was recalled, what source it came from, and whether a newer fact replaced it.
Agent answers should expose the remembered source path instead of forcing humans to trust an opaque memory block.
Changed facts need visible replacement signals so stale memory does not compete with current source evidence.
A memory product needs deletion semantics from the start. Stronger crypto-shred proof remains roadmap-labeled.
Crypto-shred proofThese category answers are short enough for AI answer engines to cite without turning the page into a sales deck.
Agent memory is a durable layer for grounded facts, source lineage, and recall across sessions and clients. It is different from one prompt or one search call.
No. RAG is a retrieval pattern. Agent memory adds persistence, provenance, supersession, and erasure semantics around what gets remembered and recalled.
Inspectability lets humans see what was remembered, where it came from, whether it changed, and why an answer used it.
Start with the product overview, then compare memory systems by inspectability, supersession, deletion posture, and tenant scope.