dont-repeat

Stop AI coding agents from making the same mistake twice.

Set up once. Work normally. Agents learn.

You should not type every failure by hand. After init, agents are instructed to self-log mistakes, and Claude Code also gets hooks that capture failed tools automatically.

Memory is per project — switch from Claude Opus → Codex → agy anytime; they all share the same lessons.

60-second setup

npm install -g dont-repeat
cd your-project
dont-repeat init

That’s it. Open Claude Code, Codex, Cursor, Gemini, OpenCode, agy, Hermes, …

dont-repeat guide    # beginner walkthrough
dont-repeat doctor   # optional health check

No cloud. No account. No API key.

How automatic is it?

| Layer | What happens | You do | |--------|----------------|--------| | Read memory | Agents load .agent-memory/MEMORY.md via instruction files | Nothing | | Self-log (all agents) | Rules say: on failure, call memory_log or run dont-repeat log | Nothing (agent should) | | MCP (Claude / Cursor) | init writes MCP config for silent memory_log | Nothing | | Claude hooks | Failed tools → auto-log; stop → best-effort distill | Nothing | | Manual log | Optional backup if a model forgets | Only if you want |

Honest note: Auto-log is as strong as we can make it (rules + MCP + hooks). Most agents follow it; a few may skip. Manual log remains a backup — not the main path.

Cross-model: If Claude logs a failure, Codex in the same repo sees it next session.

What problem this solves

Every new session, agents:

• retry fixes that already failed
• re-discover the same gotcha
• burn tokens re-learning your project

dont-repeat keeps a small shared notebook of failures, decisions, and working commands.

Install

Need: Node.js 18+

npm install -g dont-repeat

Or: npx dont-repeat@latest init

From source:

git clone https://github.com/xCaptaiN09/dont-repeat.git
cd dont-repeat && npm install && npm run build && npm link

Everyday use

Default (recommended)

1. dont-repeat init once
2. Use your coding agent as usual
3. Optionally dont-repeat status later

Optional manual commands

dont-repeat status
dont-repeat list
dont-repeat search auth
dont-repeat doctor

# only if you want to add a lesson yourself
dont-repeat log failure "do not use jest for e2e — use playwright"

Distill notes (optional)

dont-repeat distill notes.txt --apply

How it works

your-project/
├── .agent-memory/
│   ├── store.json           # database
│   ├── MEMORY.md            # what agents read (~600 tokens budget)
│   └── HOW_TO_CONNECT.md    # any-CLI fallback
├── CLAUDE.md / AGENTS.md / …  # “read memory + self-log”
├── .mcp.json                  # Claude MCP (auto)
└── .cursor/mcp.json           # Cursor MCP (auto)

1. Lessons land in the store (agent, hook, or you)
2. Ranked into MEMORY.md under a token budget
3. Next session / next model reads the same file

Supported agents

| Agent | Flag | Automatic features | |-------|------|--------------------| | Claude Code | claude | Rules + MCP + hooks (tool-fail auto-log, stop distill) | | Cursor | cursor | Rules + MCP | | Codex | codex | AGENTS.md self-log rules | | Gemini CLI | gemini | GEMINI.md | | OpenCode | opencode | AGENTS.md (no forced opencode.json) | | Antigravity (agy) | agy | AGENTS.md + GEMINI.md | | Hermes | hermes | HERMES.md + AGENTS.md | | Kimi / Qwen / OpenClaw / ZCode | kimi qwen openclaw zcode | Instruction pointers | | Aider / Windsurf / Copilot | aider windsurf copilot | Rules / AGENTS.md | | Anything else | generic | MEMORY.md + HOW_TO_CONNECT |

dont-repeat init --agents all
dont-repeat init --agents claude,codex,agy
dont-repeat init --agents antigravity   # alias → agy

Models (MiMo, MiniMax, …): use them inside a host CLI — memory follows the host.

Unsupported CLI?

Still works:

1. dont-repeat init
2. Point the tool at .agent-memory/MEMORY.md
3. Or add the self-log one-liner from HOW_TO_CONNECT.md
4. Or wire MCP: dont-repeat mcp

Command reference

| Command | Purpose | |---------|---------| | init | Main setup — do this once | | guide | Beginner walkthrough | | status / doctor / list / search | Inspect | | log | Optional manual lesson | | distill | Extract from notes/transcript | | mcp | MCP server | | render / budget / forget / path | Advanced |

dont-repeat --help
dont-repeat help init

Privacy

• Local only — no cloud, no telemetry
• .agent-memory/ gitignored by default

FAQ

Do I have to log every failure myself?
No. That’s the point of v0.6 — agents + Claude hooks handle it. Manual log is backup.

I switched from Opus to Codex — will Codex know?
Yes, if the failure was logged into this project’s memory (by agent, hook, or you).

Is auto-log 100% for every model?
As strong as rules + tools + hooks allow. Claude Code is strongest. Others depend on following instructions.

API key?
Not required.

Credits

Muhammed Dilshad A — @xCaptaiN09

• npm: https://www.npmjs.com/package/dont-repeat
• GitHub: https://github.com/xCaptaiN09/dont-repeat

License

MIT © 2026 Muhammed Dilshad A (xCaptaiN09)