Monorepo, only the memory that matters

Small teamsFrontend and backend developers sharing one repo

Our room filled with hundreds of notes. A frontend prompt shouldn't get database tuning notes taking up its context.

How do I keep irrelevant notes out of my AI agent's context in a shared monorepo?

Let ArcRoom's recall pick what the agent sees: it scores every candidate memory against the prompt and packs the best ones under one size budget. A CSS question gets UI decisions and the web code map; a migration question gets database notes. Scoring can miss, and if a needed memory isn't picked, the agent can still search with find_task.

Also asked as: Our frontend prompts are getting database notes they don't need. How do I stop that? · How do I give an AI agent only the memory that fits the task? · Can shared AI memory stay small when there are hundreds of notes?

The work

Recall scores every candidate memory against the prompt.

ArcRoom keeps
  • Code map
  • Decision
  • Recall
The next agent gets

Recall (budget 1.2k tokens): 3 of 214 memories

Step by step

  1. Recall scores every candidate memory against the prompt.
  2. It packs the best ones under one size budget.
  3. A CSS question gets UI decisions and the web code map; a migration question gets database notes.

What the agent sees

Recall (budget 1.2k tokens): 3 of 214 memories
• Decision: Tailwind tokens only, no inline hex
• Code map: apps/web/components/ui
• Answer: dark-mode toggle lives in ThemeProvider

Compared with today's tools

Folder-level CLAUDE.md / Cursor rules with globs

Put a CLAUDE.md in each package, or scope Cursor rules by file glob, so only nearby rules load.

Setup
1–2 hours
Ongoing
keep each file current

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

Folder scoping is precise and predictable; ArcRoom scores by what the prompt asks, which can miss, but needs no files per folder.

Where it stops

Scoring can miss. If a needed memory isn't picked, the agent can still search with find_task.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in small teams