Zenrei（前例）

"That proposal? There's a precedent." — Decision memory for AI-driven development.

AI agents don't remember past discussions. They keep re-proposing designs you rejected weeks ago. Zenrei records your decisions — accepted, rejected, deferred — as Markdown in your repos, and serves them to every AI agent (Claude Code / Cursor / CI) over MCP. When an agent re-proposes a rejected design, check_proposal stops it:

zenrei › check_proposal
⚠ Conflicting precedent found
  [rejected] Remove the approval flow from schema chat
  (2026-07-08)
  Reason: the pre-write review layer is our moat.

• git-native — decisions live in your git. Zenrei indexes and serves; it never owns your data
• Humans decide — agents can only file drafts (pending); only a human can accept or reject
• Deterministic & light — no LLM calls, no API key, works offline (Japanese-aware bigram BM25). With 1,000 decisions, responses stay ~1,800 chars / 44ms. Your prompts never bloat

Quick start

claude mcp add zenrei --scope user -- npx -y zenrei --root /path/to/repo1 --root /path/to/repo2

For English tool responses, append --lang en (default is Japanese; or set ZENREI_LANG=en).

Then add three lines to each repo's CLAUDE.md (agent instructions):

## Decision memory (zenrei)
- At session start, call get_context (scope: this repo) to load decided matters
- Before proposing designs or policies, call check_proposal
- File new decision points with record_decision (humans approve)

Writing a decision

decisions/2026-09-07-example.md:

---
status: rejected            # accepted | rejected | deferred | pending | superseded
date: 2026-09-07
scope: [my-repo]
tags: [design]
---

# Adopt Redis for the cache layer

## Decision
Rejected. In-memory is sufficient at current traffic.

## Reason
Not worth the operational cost.

## Revisit when
p95 latency exceeds 200ms.

Agents can only create pending drafts. Flipping status: to accepted is a human's job.

Tools (fixed at five)

check_proposal / search_decisions / record_decision / list_pending / get_context

Links

• How it works: https://zenrei.karakurio.com
• Operations guide: https://zenrei.karakurio.com/guide
• Format spec: https://github.com/pikonori/zenrei/blob/main/docs/decision-format.md

日本語

Zenrei（前例）は、AI駆動開発のための意思決定メモリです。採用・却下・保留の決定をリポジトリ内のMarkdown（decisions/*.md）で管理し、MCP経由で全AIエージェントに配ります。エージェントが却下済みの設計を蒸し返したら check_proposal が止めます。

• 決定はあなたのgitの中（データを人質に取りません）／決定するのは人間（AIは起票まで）
• 照合は決定論的（LLM不使用・APIキー不要・オフライン動作）。決定が増えてもプロンプトは太りません

導入は上記Quick startのコマンドそのまま。仕組みの解説は https://zenrei.karakurio.com 、使い方は https://zenrei.karakurio.com/guide を参照してください。

License

MIT