Decisions you made weeks ago
Solo buildersSolo founder who changed their mind once and doesn't want to again
Three weeks ago we decided permissions live in Postgres row-level security, not in API code. New sessions kept suggesting a middleware permission check, and twice an agent half-built one before we noticed.
How do I make my AI agent remember an architecture decision I made weeks ago?
Record the decision once with ArcRoom: arcroom decide "permissions in RLS only" --why "one place to audit". Before each prompt, recall picks the decisions that fit the request and puts them in front of the agent, so it builds inside the decision instead of arguing with it. Recall works under a size budget, so a decision worded far from the request may not be picked.
Also asked as: My agent keeps building something I decided against. How do I stop it? · How do I give Claude Code a list of project decisions it must follow? · Can an AI agent remember why I chose one design over another?
Record the decision once: arcroom decide "permissions in RLS only" --why "one place to audit".
- Decision
- Recall
Team decision (Oct 9): permissions live in RLS only, never in API code.
Step by step
- Record the decision once: arcroom decide "permissions in RLS only" --why "one place to audit".
- Before each prompt, recall picks the decisions that fit the request and puts them in front of the agent.
- The agent builds inside the decision instead of arguing with it.
What the agent sees
Team decision (Oct 9): permissions live in RLS only, never in API code.
Why: one place to audit.Compared with today's tools
CLAUDE.md
Add a line to CLAUDE.md: permissions live in Postgres row-level security, no middleware checks.
- Setup
- 5 min
- Ongoing
- edit the file per decision, prune it as it grows
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
For a handful of decisions CLAUDE.md is as good and always loaded; it costs more as it grows, and Codex or Cursor read their own files.
Where it stops
Recall chooses what fits the prompt under a size budget. A decision worded far from the request may not be picked.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.