The answer that went stale
Small teamsTeam with fast-moving code
An agent explained how rate limiting works. A week later someone rewrote the limiter, and nobody wanted the old answer handed out.
How do I stop AI agents from handing out old answers after the code they describe changed?
Save answers through ArcRoom, which stores the git hashes of the files each answer relied on. When those files change, the answer is no longer reused as-is, and the next agent redoes the work and saves a fresh answer. Freshness covers files in git: facts about outside systems, like an API's limits, need an expiry rule.
Also asked as: Someone rewrote our code and the agent still gives the old explanation. How do I prevent that? · How do I expire saved AI answers when the code changes? · Can AI agent memory tell when it's out of date?
Each saved answer stores the git hashes of the files it relied on.
- Reused answer
- Freshness check
find_task → similar, but stale: lib/ratelimit.ts changed since this was saved.
Step by step
- Each saved answer stores the git hashes of the files it relied on.
- When those files change, the answer is no longer reused as-is.
- The next agent redoes the work and saves a fresh answer.
What the agent sees
find_task → similar, but stale: lib/ratelimit.ts changed since this was saved.
Redo the lookup and save the new answer.Compared with today's tools
Docs next to the code, checked in review
Keep the rate-limit explanation in docs/ beside the code and ask for doc updates in PR review.
- Setup
- 1 hour
- Ongoing
- reviewers must remember to ask
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min per person (npx arcroom init)
- Ongoing
- None
Review only catches what reviewers notice; ArcRoom marks the answer stale on its own when a file it relied on changes in git.
Where it stops
Freshness covers files in git. Facts about outside systems (an API's limits) need an expiry rule.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.