Persistent storage that holds an AI agent's long-term memory across conversations and sessions—facts about users, preferences, prior decisions, ongoing projects. Memory stores can be unstructured (vector databases of conversation summaries), structured (key-value records keyed by user ID and topic), or hybrid. Effective memory stores include retrieval logic to surface only relevant memories on each turn, write logic to update or contradict prior memories, and decay/expiry policies to prevent stale data from corrupting future responses.
Frequently asked questions
How is a memory store different from a knowledge base?
A knowledge base is shared organizational content—help articles, policies, product specs—generally written by humans and consumed by agents. A memory store is per-user (or per-account) state generated by the agent itself: 'this customer prefers email,' 'this user already tried solution X,' 'this contract was reviewed last Tuesday.' KBs are about the company's knowledge; memory stores are about the relationship and history with each user.
How do you prevent a memory store from accumulating wrong facts?
Three patterns: (1) confidence thresholds—only commit facts when the model is highly confident, (2) contradiction handling—when new evidence contradicts an old memory, update or invalidate it explicitly rather than silently appending, and (3) human review for high-stakes facts. Some teams also run a periodic 'memory garbage collection' agent that audits the store and prunes outdated entries.