The question you ask every week
Solo buildersSolo developer on a codebase older than a month
Every few days we asked an agent some version of 'how does auth work in this app?'. Each time it opened the same six files and wrote the same answer, about 25k tokens each time.
How do I stop my AI agent from re-reading the same files to answer the same question every week?
With ArcRoom, the agent checks find_task before answering and saves its answer with complete_task, along with the git hashes of the files it read. When the question comes back in other words, find_task returns the saved answer in seconds. If any of those files changed, the answer counts as stale and the agent does the work again. This only helps when the agent actually calls find_task.
Also asked as: Why does my agent spend 25k tokens explaining the same part of my code every week? · How do I reuse an AI agent's answer about my codebase in a later session? · Can an AI agent remember answers it already gave about my code?
The first time, the agent checks find_task, finds nothing, claims the question, answers it and saves the answer with complete_task.
- Reused answer
- Claim
- Saved answer
- Freshness check
find_task → match: reuse (0.91)
Step by step
- The first time, the agent checks find_task, finds nothing, claims the question, answers it and saves the answer with complete_task.
- The answer is stored with the git hashes of the files it read.
- Next week a new session asks the same thing in other words. find_task returns a reuse match, and the agent answers from it in seconds.
- If any of those files changed, the answer is treated as stale and the agent does the work again.
What the agent sees
find_task → match: reuse (0.91)
"Auth: Supabase JWT checked in middleware.ts, roles from profiles.role, RLS on every table…"
files unchanged since saved · saved 24k tokensCompared with today's tools
A docs file linked from CLAUDE.md
Save the agent's auth answer as docs/auth.md and point CLAUDE.md at it, so later sessions read the doc instead of the code.
- Setup
- 15 min
- Ongoing
- update the doc by hand when auth changes
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 (npx arcroom init)
- Ongoing
- None
A good doc is nearly as cheap to reuse; nothing tells you when it goes stale, while ArcRoom checks the git hashes of the files the answer used.
Where it stops
Only helps when the agent actually calls find_task. Answers about fast-changing files go stale quickly and are redone.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.