How AI agents remember.

AI agents remember in two ways. Short-term memory is the context window: everything the model can see in the current session, gone when the session ends. Long-term memory is anything saved outside the model, such as a notes file, a database or a search index, that the agent reads back in a later session.

The model itself does not change while you talk to it. Everything that feels like AI agent memory is saved text that gets put back in front of the model. Once you see it that way, you can decide what gets saved, check it, and keep the wrong things out.

01Context, memory and state: three different things

These three words get mixed up, and the difference matters when something goes wrong.

• Context is what the model reads right now: your instructions, the conversation, files and tool results. It has a hard size limit, and models get less reliable as it fills. Our guide to context engineering covers how to budget it.

• Memory is what survives after the session: saved notes the agent loads next time.

• State is where a task stands: which steps are done, which files changed, what comes next. Agents that work for hours keep state in a progress file, so a fresh session can pick up the work.

02The four kinds of agent memory

A framework from Princeton researchers, called CoALA, splits agent memory into working memory plus three long-term kinds. The labels are useful because each kind needs different care.

The paper adds a warning worth remembering. Letting an agent rewrite its own procedures is “significantly riskier” than letting it save facts or experiences, because it can introduce bugs or work around its designers’ intent. In practice: let an agent take notes, but review any change to its rules.

03Where long-term memory lives

Long-term memory is just storage. The three common kinds trade simplicity for scale.

Start with files. You can read them, edit them and keep them in version control, and most agents handle them well. Move to vector search when memory outgrows what you can load, which is the same machinery as RAG. Embeddings explains how the “search by meaning” part works.

04AI agent memory in ChatGPT, Claude and Gemini

As of September 2026, all three big assistants remember across chats. Each lets you view, edit or switch off what it keeps.

One detail stands out. OpenAI’s help page says the memory summary does not include everything ChatGPT remembers. Deleting one entry is not always the same as forgetting. Memory is also separate from model training, which has its own switch. Our AI privacy guide covers the training settings.

Coding agents use files you can open. Claude Code loads your CLAUDE.md instructions at the start of every session. Its auto memory is Claude’s own notes, one folder per repository, with a MEMORY.md index whose first 200 lines or 25KB load each time. All of it is plain Markdown you can read, edit or delete.

05How to give an agent memory safely

1. 01

   Start with one notes file per project

   A single Markdown file, kept in the project and in version control, beats a clever memory system you cannot see. You can read every line and undo any change.

2. 02

   Separate rules from notes

   Rules you write, such as CLAUDE.md or AGENTS.md, say how to work. Notes the agent writes record what it learned. The agent may add notes freely, but changes to rules go through you.

3. 03

   Give each entry a date and a source

   “2026-09-14: staging uses port 8080 (from deploy logs)” can be checked and retired. “Staging uses 8080” cannot.

4. 04

   Review what was saved

   Skim the diff of the notes file after important sessions. Wrong notes are worse than none, because the agent trusts them.

5. 05

   Prune on a schedule

   Delete entries that are stale or no longer used. Anthropic’s memory tool docs recommend expiring files that have not been accessed in a long time.

Here is a notes file that stays readable as it grows:

# Project notes (the agent adds here; I review weekly)

## Decisions
- 2026-09-10: Payments go through the /billing service, never direct API calls (decided in review)

## Facts
- 2026-09-14: Staging uses port 8080 (from deploy logs)

## Open problems
- 2026-09-18: Login test is flaky on CI, cause unknown

## Do not store here
Passwords, API keys, customer data, anything copied from web pages or emails

And a prompt that ends a session with a clean memory update:

Before we stop, update NOTES.md.
Add only what a future session would need: decisions we made and why, facts you confirmed, and problems still open. Date each entry and say where the fact came from.
Remove or correct any entry that today's work proved wrong.
Do not save secrets, personal data, one-off instructions, or anything you read in web pages or emails.
Show me the diff before you save.

06What to keep out of agent memory

Memory turns a one-time mistake into a permanent one. That makes a few things dangerous to store.

• Secrets. Passwords, API keys and tokens get replayed into every future session, and into every place those sessions send data.

• Other people’s personal data. Customer records, health details and anything under an NDA belong in systems with access controls, not in an agent’s notes.

• Anything from untrusted content. A web page or email can carry instructions written for the agent. Saved to memory, they run again in every future session.

• Guesses. An unverified note hardens into a “fact” the agent repeats with confidence.

• One-off instructions. “Skip the tests this time” should never become a standing rule.

07FAQ

Next, learn how RAG answers from your own documents, or see what context engineering changes about everything an agent reads.