Memory Store — AI Glossary

Reference details

Related terms

• Working Memory (Agents)
• Long-Term Memory (Agents)
• AI Agent

How it works

1. The store is chosen to match the access pattern: vector search for similarity recall, key-value for exact lookups by user or session, relational where entries must be queried and joined.
2. Each entry carries scope — working, episodic, semantic or procedural — because retrieving an event when a fact was wanted is a category error the store cannot catch.
3. Writes are mediated by a policy rather than appended freely, since the store has no view on what deserves to persist.
4. A supersede or merge rule resolves contradictions on write, which is the single most consequential piece of the design and the one a database swap does not provide.
5. Retrieval filters by scope and recency before similarity, so a stale entry does not outrank a current one on phrasing alone.

Example: The part a database migration does not fix

A team moves from one vector database to another, hoping to fix an agent that keeps recalling wrong preferences. Recall latency improves and the wrong answers continue, because both stores faithfully return what was written and the problem was that a changed preference was appended rather than superseding the old one. Two contradictory entries were present throughout and the retriever ranked on phrasing. The store was never the variable; the write policy was.

Commonly confused with

The store is where memories live; memory is the behaviour of using them. Swapping a vector database for a different one changes the store and nothing about the hard part, which is the policy — what gets written, what gets overwritten when a fact changes, and what is allowed to be forgotten. A store with no forgetting policy grows until retrieval quality collapses.

When to use it

Reach for it when:

• Choosing infrastructure for agent memory, where the access pattern should drive the choice.
• Separating scopes so facts, events and learned procedures are not retrieved interchangeably.
• Reasoning about where a memory bug lives — the policy or the retrieval.

Reach for something else when:

• As the answer to a memory quality problem, which is usually the write policy.
• Where the context window already holds everything the task needs.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Memory Store — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path