Context and memory: what your agent knows, and when

Every agent session starts with an empty head. What the agent knows about your project comes from four places: files it loads at the start, files it pulls in as it works, notes it wrote in earlier sessions, and the conversation itself. Put each piece of knowledge in the right place and the agent follows your conventions without reminders. Put it in the wrong place and it gets ignored, lost at compaction, or paid for on every request.

The context window is everything the model reads on one request: Claude Code's system prompt, your instruction files, the memory index, tool names, skill descriptions, the conversation so far, every file it read and every tool result. It has a fixed size. When it fills up, Claude Code compacts: it replaces the older conversation with a summary and keeps going.

Memory is what survives between sessions. Claude Code has two kinds:

Both are context, not configuration. The agent reads them and tries to follow them, with no guarantee. A rule that must hold every time belongs in a permission rule or a hook, which the runtime enforces whatever the model decides.

In Claude Code, each mechanism enters the context at a different moment and costs a different amount:

Decide where a piece of knowledge goes by how often the agent needs it and whether it must be enforced:

Add to CLAUDE.md when Claude makes the same mistake twice, when a code review catches something it should have known, or when you type the same correction you typed last session. Leave out what the agent can read from the code: directory trees, dependency lists and architecture overviews cost tokens on every request and go stale.

AGENTS.md is the cross-tool file. It is plain Markdown with no required fields, contributed by OpenAI and now stewarded by the Agentic AI Foundation under the Linux Foundation, next to MCP. In a monorepo the nearest file in the directory tree wins, and your own prompt overrides any file. Support today:

App memory is a separate layer. Claude, ChatGPT and Codex can remember things about you across chats. That helps with personal preferences, but nobody on your team can review it, so team rules belong in files in the repository.