Memory

Helps:

• Standing preferences: language, tone, format, your role.
• Project facts that never change: tool names, folder locations, team members.
• Lessons learned: "do not use library X, it broke last time".

Hurts:

• Facts that go stale (prices, deadlines, who is on the team).
• Mixing personal and client work in one memory.
• Assuming the assistant remembers something it never saved. If it matters, write it down where you can see it.

Chat apps: open Settings → Memory (or Personalization). You can see every saved note, delete single ones or turn memory off. Tell the assistant directly: "remember that I report in euros".

Claude Code and similar agents: create a CLAUDE.md (or AGENTS.md) file in the project. Put the build and test commands, the house rules and the gotchas there. It is read at the start of every session, and you can edit it like any file.

Agents you build yourself: store memory as a small document or a database table the agent reads at the start and writes to at the end. Keep it short; it is sent as tokens on every run.

• Review saved memories monthly and delete what is wrong or old.
• Write memories as facts with a date: "Since Nov 2026 invoices go through tool X".
• Keep client or project memory in that project's own file, not in your global settings.
• When a session goes wrong because of a forgotten rule, add the rule to the memory file rather than repeating it in chat.
• Treat memory as hints, not truth. For anything with consequences, look at the real document.

• Memory is not a knowledge base. It is a few hundred words of preferences, not your document archive.
• A wrong memory repeats itself in every future conversation until you delete it.
• Shared accounts share memory; a colleague's preferences become yours.
• Memory can contain personal data. Check what your organization allows before enabling it for client work.

Memory, context and knowledge base are three sizes of the same question: what does the model see right now?