Introduction

Stop teaching your agent the same lesson twice.

Oh My Experience (OME) helps your agent understand your way of working over time. It is a local-first experience recall layer for AI coding agents. It turns real Codex and Claude corrections, rework, and delivery lessons into reviewed experience cards, then recalls the most relevant reminder when a similar task appears again.

OME is not about writing more rules. It is about timing: showing the right lesson while the agent can still change its next action.

Why Developers Use It

AI coding agents are getting better at writing code, but they still lack durable judgment for your way of working.

OME helps you:

• Keep AGENTS.md and CLAUDE.md small.
• Stop losing hard-won execution lessons in chat history.
• Help agents remember the right skill, check, or release gate.
• Review every lesson before it becomes active.
• Stay local by default.

Each Layer Has A Job

AI agents already have many ways to keep context, but each layer solves a different problem.

• AGENTS.md and CLAUDE.md are best as always-on entries: norms, maps, capability indexes, and collaboration boundaries.
• Skills are best for repeatable workflows such as validation, release checks, or external-system operations.
• Memory is best for facts, preferences, and long-term background that help the agent understand you and the project.
• Rules are best for stable, pre-declared behavior constraints that apply by scope.

OME adds a different layer: execution experience.

Execution experience usually comes from a judgment formed during real work: why a step mattered last time, when it should matter again, which similar cases should not trigger it, and what reminder the agent should see before acting.

It is not always-on project guidance, and it is not a tool manual. It is recalled lightly when the current task matches, so rule files stay small and lessons do not get lost.

See Examples for a concrete /goal case that shows the matched experience card and the exact context mounted into the agent prompt.

How It Works

real work -> reflect scan -> candidates -> human review -> active cards
-> prompt-time recall -> match stats -> ongoing maintenance

1. Import sessions from local Codex history, or optionally from Spool.
2. Run a reflect scan so the agent finds places where you corrected it.
3. Review candidates one by one: approve, reject, merge, or rewrite.
4. Activate cards only after they are confirmed.
5. Recall at prompt time when the hook matches active cards to the current task.
6. Maintain the library with stats and evals for coverage, stale cards, and noisy matches.

What You Get

• A local Markdown library of reviewed experience cards.
• Optional project libraries at <project-root>/.oh-my-experience/.
• Codex and Claude hooks for prompt-time recall.
• Project-aware matching from one local hook.
• Explicit ignore criteria for noisy words such as "goal", "review", and "release".
• Explainable recall: matched cards, scores, reasons, and rendered context.
• A Markdown-first review loop for candidates before they become active.
• Isolated evaluation so retrieval changes do not pollute your real library.

Local And Reviewed By Default

The hook path does not call an LLM, use the network, or require an API key. If recall fails, it fails open and returns empty context so your agent still runs.

By default, hook events store prompt hashes and task summaries, not raw prompts. Candidates created by reflect scans never become active automatically. Nothing bypasses human review to enter prompt-time recall.

Next

Start with the Quickstart to install OME and verify your first recall. Use Global And Project Libraries if you want repository-level cards. Then read Reflect and Review to turn real corrections into reusable experience.