Chapter 19: Memory & Instructions — let the AI remember you and follow your rules
Normal AI chat "forgets" every session: close the conversation and it no longer remembers who you are or what you like.
Memory gives the AI a long-term notebook, which together with AGENTS.md project instructions keeps your preferences and rules across sessions.
Reading time: about 10 minutes.
19.1 What is "memory"? — the AI's long-term notebook
- You tell it "write docs in Chinese", "create a TODO before starting tasks", "update the README changelog before pushing a plugin";
- It writes down these explicit, durable preferences;
- In every future conversation, it automatically looks up relevant entries and acts according to your habits.
How to make the AI remember something?
Just say: "Remember: from now on, write all code comments in Chinese." The AI saves it to memory.
19.2 Memory usage essentials
- What to store: explicit preferences, working styles, communication habits (e.g. "reply in Simplified Chinese");
- What NOT to store: passwords/keys, temporary task states, one-off context (they leak or go stale);
- On-demand injection: the AI only uses memory relevant to the current task — it won't recite the whole notebook;
- Always editable: you can say "change the memory entry 'reply in Chinese' to 'reply in English'" or "delete that memory";
- Management entry: view and manage the AI's saved memory in Settings → Memory;
- Privacy: memory stays on your device and travels with backups (see Chapter 20).
19.3 Who's in charge? — priority of instructions
text
Your direct request > current conversation > project instruction files (AGENTS.md) > memory > examples- Project instructions apply per directory tree: deeper directories override shallower ones (how to write AGENTS.md: Chapter 12, 12.6);
- No instruction (including AGENTS.md and memory) may demand: leaking secrets, bypassing approvals, or unrelated external operations.
End of chapter · Next: Backup & Migration — keep your data safe