Guide
Stop Claude Code and Codex asking what you already know
Coding agents stop to ask what was decided, how you like things done and who someone is, and most of the answers already exist somewhere. Here's how to let Claude Code, Codex and Cursor look them up in Cachejay first, in about ten minutes.
Why agents ask
An agent knows your code, its instructions and whatever you typed. It doesn't know what happened outside the editor: the pricing agreed on Tuesday's call, the rule you gave a different agent last month, who “Dana from Acme” is. So it asks, or it guesses.
Both cost you. A question means the agent waits until you're back. A guess means redoing the work. The fix is to give the agent somewhere to look, and tell it to look there before asking. Why this matters more as agents run longer: when your agent needs you and you're not there.
1. Connect your agents
- Install Cachejay and sign in. It starts keeping your calls, projects, decisions and to-dos as you work.
- In the menu bar, choose Connect AI Apps, and pick Claude Code, Codex, Cursor, Antigravity or any other agent you use.
- Restart the agent so it loads the new tools.
For Claude Code, Codex, Antigravity and Cursor, connecting also installs a Cachejay skill, which already tells the agent to read before asking. Step 2 makes it explicit and covers the cases the skill doesn't. Agent-specific notes: Claude Code · Codex · Cursor.
2. Add the rules
Paste this into CLAUDE.md (Claude Code) or AGENTS.md (Codex, Cursor and most others). In a repository it applies to that project. In ~/.claude/CLAUDE.md or ~/.codex/AGENTS.md it applies everywhere. Not sure which file you need? AGENTS.md vs CLAUDE.md.
## Before you ask me
- Look it up in Cachejay first:
- what was decided, said on a call, or done before: search_memory
- how I like things done: about_me
- who someone is: people_now
- where a project stands and what's open: project_brief, todos
- Say what you found and where it came from (the call, the date).
- Still ask me before spending money, sending anything to another person,
deleting things, or making a choice nobody has made yet.
- If you're blocked on one question, carry on with the rest and list
your questions at the end.The last two rules matter as much as the first. Asking for the source lets you check the answer in a few seconds. Listing the off-limits cases keeps “look it up” from turning into “decide for me”. And “carry on with the rest” means one open question doesn't stall the whole job.
3. Test it with three questions
Pick things you know the answer to, so you can tell whether it found the right one:
- A decision: “What did we decide about [topic] on [project]? Give the source and date.”
- A preference: “How do I like pull requests written?”
- A person: “Who is [name], and what's open with them?”
› what did we decide about the export format? source and date ⎿ cachejay · search_memory CSV with one row per day. Decided on the Acme call on Monday, October 12: "Let's not do Excel, CSV is fine as long as the dates sort." Nothing later changes it.
Check the source, and check whether something later replaced it. If an answer is wrong because your words changed meaning over time, tell the agent. It can correct Cachejay's record so the next agent doesn't repeat the mistake.
4. Let them report back
The skill also has the agent tell Cachejay when it finishes something: what's done, what was decided, what's next. To-dos the work finished get ticked, and the project's status is rewritten from it. Your Claude Code and Codex sessions are kept too, so an answer you gave one agent can be found by the next instead of asked again.
When it still asks
- It asks something Cachejay should know. Ask it to search first, and see what comes back. If the call wasn't recorded or the app is in your private apps, the answer isn't there.
- It finds an old answer. Ask for dates on everything. Decisions change, and the newest one wins.
- It's a real decision. Good. That's the question it should ask. More on drawing the line: what an AI agent should ask you.