hotai.news
← Back to AI WikiAI Wiki / Core concepts

What is agent memory? Conversation state, persistent data, and model parameters

Understand agent memory through a project assistant’s preferences, distinguishing short-term state, persistent storage, retrieval, provenance, updates, and deletion.

HotAI.news · Sources checked with CodexPublished on 2026-10-20Sources checked on 2026-10-20
Our take

Understand the concept

Agent memory retains information for later use through conversation state or cross-session storage. Common implementations retrieve external data into context rather than changing model parameters. Accuracy depends on writing and updating rules. Stored information is not forever valid and does not authorize cross-user access.

01

Distinguish short-term state from persistent information

The current file can be session state, while an explicit preference for a date format can persist across sessions. Lifetimes and update rules differ. Visible chat content alone does not establish retrievable long-term memory.

LangChain distinguishes thread-local state from cross-thread memory, and Anthropic discusses external notes for task continuity. These are engineering mechanisms, not full simulations of human memory. No personal memory store was accessed or product memory tested here.

02

Store provenance and conditions with remembered content

Discussing a Friday trip should not automatically become a lasting preference. Keep model guesses distinct from confirmed information. Record source, time, and project scope where relevant so an old decision is not applied to unrelated work.

Retrieval can return outdated versions after updates. Handle replacement or invalidation and resolve conflicts against newer reliable information. Vector storage does not automatically know what is obsolete; similarity finds relevance rather than truth or version precedence.

03

Retrieval and access control govern practical use

Persistence does not require sending everything on every turn. Retrieve relevant authorized information with user isolation. Instructions in stored notes retain their source authority; being remembered does not grant new execution authorization.

Test one explicit preference: verify retrieval in a new session, then update or delete it and check for stale reuse. These are validation steps without writing business state. Traceable behavior, not the statement “I remembered,” establishes functionality.

Checked sources

These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.

Back to AI Wiki ↗