Oh My ExperienceStop teaching your agent the same lesson twice.

Turn real coding corrections into approved experience cards that surface at prompt time — only when the task actually needs them.

Try in 2 minutes

See real recall

Conditional recall, not always-on bloat

Hook decomposes prompts into task envelopes with deterministic BM25-like matching. Cards inject only when triggered — otherwise nothing.

Nothing enters without draft approval

Cards follow candidate → draft → active → archived. AI-generated drafts wait for your approval first. Only active cards are recalled.

One library, Codex and Claude

Same cards, same retrieval engine, shared across providers. Install each agent hook and skill once.

Local-first, private by default

Everything on your machine. No cloud, no accounts. The hook path calls no LLM, stores no raw prompts.

Explainable matching

Agents and debugging scripts can inspect which cards hit, why they scored, and what context would be injected.

Survives context compression

AGENTS.md injects once then gets compressed away. OME's hook re-injects on every UserPromptSubmit.

Choose Your Path

New to OME? Start with Quickstart.
Want proof first? See the /goal recall example.
Evaluating the design? Read Retrieval Engine.

A Real Recall Shape

When the user sends a real execution prompt, the installed hook gives the agent a compact candidate list with the matched card, why it matched, and the command for reading the full rule.

OME does not load every lesson all the time. It mounts the relevant candidate when the prompt looks like the workflow that needs it.