grov

Collective AI memory for engineering teams.

Website • Dashboard • Quick Start • Team Sync • Contributing

Grov captures reasoning from your Claude Code sessions and injects it into future sessions. Your AI remembers what it learned.

The Problem

Every time you start a new Claude Code session:

• Claude re-explores your codebase from scratch
• It reads the same files again
• It rediscovers patterns you've already established
• You burn tokens on redundant exploration

Measured impact: A typical task takes 10+ minutes, 7%+ token usage, and 3+ explore agents just to understand the codebase.*

The Solution

Grov captures what Claude learns and injects it back on the next session.

What Gets Captured

Real reasoning, not just file lists:

Architectural decisions, patterns, and rationale - automatically extracted.

Quick Start

npm install -g grov   # Install
grov init             # Configure (one-time)
grov proxy            # Start (keep running)

Then use Claude Code normally in another terminal. That's it.

How It Works

Session 1: Claude learns about your auth system
           ↓
        grov captures: "Auth tokens refresh in middleware/token.ts:45,
                        using 15-min window to handle long forms"
           ↓
Session 2: User asks about related feature
           ↓
        grov injects: Previous context about auth
           ↓
        Claude skips exploration, reads files directly

Commands

grov init         # Configure proxy URL (one-time)
grov proxy        # Start the proxy (required)
grov proxy-status # Show active sessions
grov status       # Show captured tasks
grov login        # Login to cloud dashboard
grov sync         # Sync memories to team dashboard
grov disable      # Disable grov

Data Storage

• Database: ~/.grov/memory.db (SQLite)
• Per-project: Context is filtered by project path
• Local by default: Memories stay on your machine unless you enable team sync

Team Sync

Share memories across your engineering team with the cloud dashboard.

grov login                    # Authenticate via GitHub
grov sync --enable --team ID  # Enable sync for a team

Once enabled, memories automatically sync to app.grov.dev where your team can:

• Browse all captured reasoning
• Search across sessions
• Invite team members
• See who learned what

Memories sync automatically when sessions complete - no manual intervention needed.

Requirements

• Node.js 18+
• Claude Code

Advanced Features

Anti-Drift Detection

Grov monitors what Claude does (not what you ask) and corrects if it drifts from your goal.

• Extracts your intent from the first prompt
• Monitors Claude's actions (file edits, commands, explorations)
• Uses Claude Haiku to score alignment (1-10)
• Injects corrections at 4 levels: nudge → correct → intervene → halt

# Test drift detection
grov drift-test "refactor the auth system" --goal "fix login bug"

Extended Cache (Experimental)

Anthropic's prompt cache expires after 5 minutes of inactivity. If you pause to think between prompts, the cache expires and must be recreated (costs more, takes longer).

grov proxy --extended-cache

What this does: Sends minimal keep-alive requests (~$0.002 each) during idle periods to preserve your cache.

Important: By using --extended-cache, you consent to Grov making API requests on your behalf to keep the cache active. These requests:

• Use your Anthropic API key
• Are sent automatically during idle periods (every ~4 minutes)
• Cost approximately $0.002 per keep-alive
• Are discarded (not added to your conversation)

This feature is disabled by default and requires explicit opt-in.

Environment Variables

# Required for drift detection and LLM extraction
export ANTHROPIC_API_KEY=sk-ant-...

# Optional
export GROV_DRIFT_MODEL=claude-sonnet-4-20250514  # Override model
export PROXY_HOST=127.0.0.1                        # Proxy host
export PROXY_PORT=8080                             # Proxy port

Without an API key, grov uses basic extraction and disables drift detection.

What Gets Stored

{
  "task": "Fix auth logout bug",
  "goal": "Prevent random user logouts",
  "files_touched": ["src/auth/session.ts", "src/middleware/token.ts"],
  "reasoning_trace": [
    "Investigated token refresh logic",
    "Found refresh window was too short",
    "Extended from 5min to 15min"
  ],
  "status": "complete"
}

What Gets Injected

VERIFIED CONTEXT FROM PREVIOUS SESSIONS:

[Task: Fix auth logout bug]
- Files: session.ts, token.ts
- Extended token refresh window from 5min to 15min
- Reason: Users were getting logged out during long forms

YOU MAY SKIP EXPLORE AGENTS for files mentioned above.

How the Proxy Works

1. grov init sets ANTHROPIC_BASE_URL=http://127.0.0.1:8080 in Claude's settings
2. grov proxy intercepts all API calls and:
   • Extracts intent from first prompt
   • Injects context from team memory
   • Tracks actions and detects drift
   • Saves reasoning when tasks complete

Roadmap

• [x] Local capture & inject
• [x] LLM-powered extraction
• [x] Local proxy with real-time monitoring
• [x] Anti-drift detection & correction
• [x] Team sync (cloud backend)
• [x] Web dashboard
• [ ] Semantic search
• [ ] VS Code extension

Contributing

1. Fork the repo and clone locally
2. Install dependencies: npm install
3. Build: npm run build
4. Test locally: node dist/cli.js --help

npm run dev              # Watch mode
node dist/cli.js init    # Test CLI

Found a bug? Open an issue.

License

Apache License 2.0 - see LICENSE file for details.