# Learning Loop

> AgentsRoom reviews what you typed to the agents of a project (the corrections you keep repeating: "no, not like that", "I already told you", "always do X first") and proposes up to 10 short rules per review.

- Area: Desktop app
- Plans: All plans
- Last checked against the product: 2026-11-12
- Web page: https://agentsroom.dev/docs/lessons

## What it does

AgentsRoom reviews what you typed to the agents of a project (the corrections you keep
repeating: "no, not like that", "I already told you", "always do X first") and proposes up to
10 short rules per review. You accept, edit or refuse each one. Accepted rules are written into
the shared project memory (two notes pinned at the top of the memory index, "Read before any task"), so every agent of the project applies them, whatever its CLI and
whatever the machine. A refused rule is never proposed again, and each accepted rule shows
whether you had to correct it again since.

## Where to find it

Desktop: **graduation cap button in the dock**, Knowledge group, next to Memory. Its badge counts the lessons of the project and turns amber when some wait for your verdict; the tooltip says how many. Since 2026-11-07 the plain count only shows while the mouse is over the dock row (or the button has keyboard focus); the amber badge stays visible at rest. The Memory window also has a **Learning Loop** button at the right of its header, marked Recommended, just left of New memory. Beta: shown only to beta testers.
Mobile: project screen > tool dock at the bottom > graduation cap button (Learning Loop screen), amber with the number of lessons waiting.

Settings > Interface elements > Dock can hide the button like any other; the Simple and Balanced interface modes fold it.

**Reminders above the composer** (desktop, since 2026-11-12). Once per launch and per project, at most one block:
- **"N lessons to validate"** when at least 5 lessons of the project were left waiting. Click it to open Learning Loop; the cross hides it until the next launch; it leaves by itself once none is waiting. The phone shows the same block in an agent's console.
- **"Turn your corrections into rules"** when the project never had a review, the daily review is off and you sent at least 50 prompts to its agents over the last two weeks. Clicking it opens Learning Loop; clicking it or its cross ("Don't suggest again") puts it out for good on every project. Settings > Composer > Smart hints > "Show dismissed hints again" shows it again.

## How to use it

1. Open Learning Loop and press **Run the review**. The first review reads the last two weeks of
   your prompts to the agents of the project; the next ones start where the last one ended.
   When the agent has finished, a "lessons to validate" block appears above its composer, on the
   desktop and on the phone: click it to go straight to Learning Loop.
2. A few minutes later, the **To validate** tab lists the proposed lessons, each with your own
   words as evidence: **Accept**, **Edit** (then "Accept this wording") or **Refuse** each one.
   The right column shows the latest review ("What matters to you", "Mistakes you correct most").
3. Turn on **Daily review**, at the bottom of the right column, to get a new review every evening.

The **Accepted** tab lists what your agents now follow, and warns when you had to correct a rule
again since you accepted it. The search bar looks through the rules, your quotes and the category
names. In a narrow window, the right column becomes a third tab, **Latest review**.

Accepted lessons can be edited or removed later (**Remove** keeps them from coming back).

## Settings

- `lessonsEngine` (project, Learning Loop window > "Who runs the review"): `cli` (default,
  "My CLI": a review agent runs on your own CLI, no AgentsRoom quota, your prompts stay with
  the provider that already received them) or `service` ("AgentsRoom service": your masked
  prompts go through the AgentsRoom server to OpenAI, after a consent line).
- `lessonsServiceDaily` (project): daily review when the engine is the AgentsRoom service.
  With "My CLI", the daily review is the project's "Lessons review" trigger (Triggers list).
- `lessonsServiceConsentAt` (project, read-only): when you consented to the service engine.

## Agent tools (MCP)

- `lessons_collect`: the review agent reads the masked prompts of the period and the analysis
  instructions.
- `lessons_submit`: the review agent hands its review back (refused with the list of errors
  when it does not match the expected shape).

## Providers

All CLIs can run the review. Prompts are read from the transcripts of Claude, Codex, Grok,
Mistral Vibe, oh-my-pi, Pi and Antigravity; for the other CLIs, from the prompts AgentsRoom
recorded when you sent them (lost after a `/clear`).

## Mobile

Present: same lists and verdicts, same search and tabs. The third tab, **Latest review**, holds the
review, **Run the review**, the daily switch and the engine: these three are asked to your desktop,
which must be open.

## Limits

- Beta: only accounts in the beta programme see it (feature flag `lessons`).

- "My CLI": unlimited on every plan, uses your CLI's own quota for a few minutes per review.
- AgentsRoom service: Free 3 reviews / month, Plus 3 / month, Pro unlimited, unlimited with your
  own OpenAI key. The desktop must be open (prompts are collected on your machine).
- Secrets are masked on your machine before any analysis; raw prompts are never stored on the
  AgentsRoom server (each lesson keeps at most two short masked excerpts as evidence).
- Up to 10 lessons per review; about 40k tokens of prompts per review, newest first.

## Common questions

- **Can AgentsRoom stop my agents from repeating the same mistakes?** Yes: run a Learning Loop review
  and accept the rules it proposes; every agent of the project reads them from the project memory.
- **Does it send my prompts to a new company?** Not by default: the review runs on your own CLI.
  Only the optional AgentsRoom service sends masked prompts to OpenAI, after you consent.
- **Does it work with Codex or Gemini, not only Claude?** Yes, lessons are shared by every CLI.
- **Will a refused lesson come back?** No, a refused or removed lesson is never proposed again.
- **Where did Learning Loop go? It was in Project memory.** It has its own button in the dock now
  (graduation cap, next to Memory) and its own window. The Memory window keeps a button to it.
- **Can an agent open it for me?** Yes: `ui_open` with the surface `lessons`.
- **Is Learning Loop on automatically for every account?** No. It is in beta (beta testers only), and
  even there nothing runs by itself: a review starts when you press **Run the review** or turn on
  **Daily review**. The app only suggests it, once, above the composer after about 50 prompts in a
  project that never had a review.
- **I keep forgetting the lessons waiting for me.** When 5 or more wait, a "lessons to validate"
  block shows above the composer once per launch; the dock button stays amber meanwhile.

## Related

- [Project Memory](https://agentsroom.dev/docs/project-memory.md): where accepted lessons are stored and read by agents.
- [Scheduled Tasks](https://agentsroom.dev/docs/scheduled-tasks.md): the daily review of the CLI engine is a trigger.
