How to give an AI agent long-term memory

An AI model only "remembers" what fits in its context window, and that window resets between runs. Long-term memory is everything you do to persist information outside that window and pull the right pieces back in when the agent needs them. This guide covers the approaches and how your choice of platform shapes them.

Why context alone is not memory

A model's context is short-term working memory: it holds the current conversation and whatever you paste in. Once the run ends, it is gone. For an agent that should recall a user's preferences next week, or build on last month's work, you need durable storage plus a way to retrieve from it — a retrieval step that injects the relevant facts back into context on demand.

The four common approaches

• Vector store (semantic memory). Embed notes, documents, or past turns as vectors and retrieve the most similar ones at query time. Best for fuzzy recall ("what did the user say they liked?"). This is the workhorse of most memory setups.
• Knowledge graph / database (structured memory). Store facts as records or triples (entities and relationships). Best for precise, queryable facts ("which project is this ticket under?").
• Files (durable notes). Let the agent read and write plain files — a scratchpad, a running log, project docs. Simple, transparent, and easy to inspect.
• Summarisation (compression). Periodically compress old history into a short summary you keep in context, so the agent retains the gist without the token cost.

Most production agents combine several: a vector store for recall, a database for structured state, and files for working notes.

Match the approach to the job

How your platform changes the work

How much you build yourself depends on where the agent runs. We record a Memory & state fact on every platform dossier, because support varies widely:

• Some frameworks ship memory primitives you assemble yourself — maximum control, more wiring.
• Some managed platforms provide persistent memory as a built-in feature, so state survives between runs without you standing up a vector database.
• Some tools leave memory entirely to you.

Compare the Memory & state facts side by side on the platforms directory, or look at platforms that score well on memory for research assistants, where recall is central.

A practical starting point

1. Start with files for anything the agent should be able to read back — it is the easiest to reason about and debug.
2. Add a vector store once you need fuzzy recall across a growing history.
3. Introduce a database or knowledge graph when you have structured facts that need exact answers.
4. Layer in summarisation when long histories start to cost too much.

Next steps

• Connect the tools your agent reads from: How to connect an MCP server.
• Keep the agent available so memory is worth accumulating: How to run an AI agent 24/7.