Repomemory

repomemory

Your codebase never forgets.

AI agents lose context every session. repomemory fixes that — one command creates a persistent, searchable knowledge base that any AI tool can read, search, and write to.

[](https://www.npmjs.com/package/repomemory) [](https://www.npmjs.com/package/repomemory) [](https://github.com/DanielGuru/repomemory/blob/main/LICENSE) [](https://github.com/DanielGuru/repomemory/actions)

npx repomemory go

---

The Problem

Every time you open a project with Claude Code, Cursor, Copilot, or any AI coding agent:

• It re-discovers your architecture from scratch
• It proposes changes that were already debated and rejected
• It re-introduces bugs that were already fixed

Your CLAUDE.md / .cursorrules helps, but it's static and gets stale.

The Solution

.context/
├── index.md              ← Quick orientation (loaded every session)
├── facts/                ← Architecture, database, deployment
├── decisions/            ← "We chose Drizzle over Prisma because..."
├── regressions/          ← "This broke before. Here's what happened."
├── preferences/          ← Your coding style — follows you across all repos
├── sessions/             ← Auto-captured AI session summaries
└── changelog/            ← Monthly git history syncs

Facts tell agents how things work. Decisions prevent re-debating. Regressions prevent re-breaking. Preferences teach agents how you code.

Quick Start

With an API key (Claude Code, terminal workflows)

npx repomemory go

One command: sets up global profile, creates .context/, configures Claude Code + Cursor, runs AI analysis, prints CLAUDE.md instructions.

With Cursor (no API key needed)

npx repomemory setup cursor

This installs everything Cursor needs:

• MCP server in ~/.cursor/mcp.json (auto-starts repomemory in every project)
• Rules in .cursor/rules/repomemory.mdc (teaches Cursor's AI to use context)
• 6 commands in .cursor/commands/ (run with / in Cursor chat)

Then in Cursor chat, type:

/repomemory-analyze

Cursor's own AI scans your repo and populates .context/ via the MCP tools. No external API key required — your Cursor subscription handles it.

Available Cursor commands:

| Command | What it does | |---------|-------------| | /repomemory-analyze | Full repo analysis — populates facts, decisions, index | | /repomemory-orient | Quick orientation at start of session | | /repomemory-search | Search the knowledge base | | /repomemory-record | Record a fact, decision, or regression | | /repomemory-session | Save a session summary | | /repomemory-status | Show context coverage |

Guided wizard

npx repomemory wizard

Walks through provider selection, tool integration, and first analysis. If no API keys are detected, offers a "None — I use Cursor" option that skips external analysis entirely.

Non-interactive (CI-safe)

npx repomemory go --yes --provider anthropic --embedding-provider gemini --max-files 80

No prompts when --yes / --defaults / --no-prompt is used.

MCP Server — Agents With Real Memory

The real power is the MCP server. When configured via repomemory setup claude, it auto-starts with Claude Code and gives agents 7 tools:

| Tool | What It Does | |------|-------------| | context_search | Hybrid keyword + semantic search across repo + global context | | context_auto_orient | One-call orientation: index, preferences, recent sessions | | context_write | Write entries with smart scope routing (preferences → global) | | context_read | Read full content, repo-first with global fallback | | context_list | Browse entries with [repo]/[global] provenance tags | | context_risk | Assess modification risk — hotspots, hidden coupling, bus factor | | context_delete | Remove stale knowledge |

Agent: "Let me orient myself in this project..."
→ context_auto_orient()
→ Returns: project overview, preferences, recent sessions, recent changes

Agent: "Let me search for context about the auth flow..."
→ context_search("authentication flow")
→ Auto-routes to facts/ category, returns compact results

Agent: "I found a race condition. Let me record this."
→ context_write(category="regressions", filename="token-refresh-race", content="...")
→ Persisted. Detects if it supersedes an existing entry.

Sessions are auto-captured on shutdown. Zero config — repomemory setup claude handles everything.

Supported Tools

| Tool | Integration | API Key Required? | |------|------------|-------------------| | Claude Code | MCP server (auto-starts) + post-commit hook | Yes (for analysis) | | Cursor | MCP server + rules + 6 slash commands | No — uses Cursor's built-in AI | | GitHub Copilot | copilot-instructions.md | Yes (for analysis) | | Windsurf | .windsurfrules | Yes (for analysis) | | Cline | .clinerules | Yes (for analysis) | | Aider | .aider.conf.yml | Yes (for analysis) | | Continue | .continue/rules/ | Yes (for analysis) |

> Cursor users: You don't need any API key. Run npx repomemory setup cursor, then use /repomemory-analyze in Cursor chat. Cursor's AI does the analysis using the MCP tools — whatever model Cursor is using (it can even switch models mid-task).

Supported Providers

| Provider | Models | Env Variable | |----------|--------|-------------| | anthropic | claude-sonnet-4-6, claude-opus-4-6 | ANTHROPIC_API_KEY | | openai | gpt-4o, o3-mini | OPENAI_API_KEY | | gemini | gemini-2.0-flash, gemini-2.5-pro | GEMINI_API_KEY / GOOGLE_API_KEY | | grok | grok-3, grok-3-mini | GROK_API_KEY / XAI_API_KEY |

Embeddings (optional, for semantic search): Gemini text-embedding-004 (free, default) or OpenAI text-embedding-3-small. Auto-detected from available API keys.

All Commands

repomemory go                            # One-command setup (add --yes for non-interactive)
repomemory wizard                        # Interactive guided setup
repomemory analyze                       # AI-powered repo analysis
repomemory analyze --merge               # Update without overwriting edits
repomemory analyze --dry-run             # Preview without API call
repomemory search                 # Search knowledge base from terminal
repomemory status                        # Coverage and freshness report
repomemory doctor                        # Diagnostics and health check
repomemory dashboard                     # Local web UI (localhost:3333)
repomemory risk                          # Hotspots, coupling, ownership analysis
repomemory risk -f src/auth.ts           # Targeted risk assessment for specific files
repomemory sync                          # Sync git history to changelog
repomemory setup                   # Configure Claude/Cursor/Copilot/etc
repomemory hook install                  # Auto-sync changelog on commits
repomemory global list                   # Manage global developer context
repomemory global export                 # Export global context as JSON

Configuration

.repomemory.json in your repo root (all fields optional):

{
  "provider": "anthropic",
  "model": "claude-sonnet-4-6",
  "embeddingProvider": "gemini",
  "hybridAlpha": 0.5,
  "maxFilesForAnalysis": 80,
  "enableGlobalContext": true
}

Custom ignorePatterns and keyFilePatterns are additive — they extend built-in defaults, not replace them.

Why Not Just CLAUDE.md?

| | CLAUDE.md | repomemory | |--|-----------|------------| | Maintenance | Manual | AI-generated + agent-maintained | | Search | Load everything | Hybrid keyword + semantic | | Cross-tool | Claude Code only | 7 tools supported | | Team knowledge | One person writes | Every AI session contributes | | Decisions | Mixed in with instructions | Structured, searchable | | Regressions | Not tracked | Prevents repeat bugs | | Sessions | Not tracked | Auto-captured on shutdown | | Freshness | Unknown | Staleness detection + auto-purge |

repomemory doesn't replace CLAUDE.md — it complements it. CLAUDE.md is for instructions and rules. .context/ holds the knowledge that grows over time.

Contributing

See [CONTRIBUTING.md](.github/CONTRIBUTING.md) for development setup, testing, and contribution guidelines.

License

MIT

---

Built for developers who are tired of AI agents forgetting everything between sessions.

Report Bug · Request Feature · npm

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DanielGuru
• Source: DanielGuru/repomemory
• License: MIT

Install and usage instructions live in the source repository linked above.