git-memory

Instant project context for Claude Code sessions — powered by your git history.

You already write great commit messages. Why not let them work for you?

The Problem

Every new Claude Code session starts cold. CC has to explore the codebase, read files, figure out what's active — burning tokens and time before it can do real work. CLAUDE.md helps, but keeping it manually updated is a chore nobody does consistently.

Worse: naive git-history tools dump everything into CLAUDE.md — build bumps, merge commits, lockfile noise — and the section bloats to 200+ lines of signal-to-noise disaster.

The Solution

git-memory indexes your git history into a lightweight SQLite database and auto-generates a lean, signal-rich project context section in your CLAUDE.md. When a new CC session starts, a SessionStart hook refreshes it automatically. CC wakes up already knowing:

• Hot files — which real code files have the most churn (lockfiles/generated files excluded)
• Active areas — which directories are seeing action
• Recent feature commits — what actually shipped: feat:, fix:, refactor: — not build bumps

All in ~50–80 tokens. Huge signal, tiny footprint.

Example Output

## Branch: main

## Hot Files (last 50 commits)
17|expo-app/src/screens/TalentProfileFormScreen.tsx
6|convex/schema.ts
4|expo-app/src/components/ui/ScheduleWidget.tsx
3|expo-app/src/screens/MyMatchesScreen.tsx

## Active Areas
82|expo-app
22|convex
8|(root)

## Recent Feature Commits
4e2a730 feat(matches): Add Tinder-style swipe gestures to new match cards
1b2a9f0 feat(matches): Add 4-tier interest levels and Tinder-style swipe cards
3233080 fix(ui): Relax Step 22 spacing, make all My Matches sections collapsible
6b77132 refactor(navigation): Route all job-board links to my-matches
967f029 fix(onboarding): Restore full 25-step flow, fix schedule dots
...

No merge commits. No chore: bump build 97. No package-lock.json churn. Just the 15 commits CC actually needs to understand what you've been building.

Install

# Clone or copy the skill
git clone https://github.com/soderalohastrom/git-memory.git ~/.skills/git-memory

# Add alias (optional but recommended)
echo 'alias git-memory="bash ~/.skills/git-memory/scripts/git-memory.sh"' >> ~/.zshrc
source ~/.zshrc

# Init any repo
cd /your/project
git-memory init

init does everything:

1. Creates .ai/memory.db with indexed commits
2. Installs a Claude Code SessionStart hook in .claude/settings.json
3. Generates the first project memory section in CLAUDE.md

Commands

Command	What it does
git-memory init	First-time setup — indexes, installs hook, writes CLAUDE.md
git-memory index	Incremental index (new commits only, run after commits)
git-memory status	Show what's indexed, branch info
git-memory context [--limit N]	Output project context (default: 15 feature commits)
git-memory refresh-claude-md	Re-generate CLAUDE.md section from current index

How It Works

• No LLM calls — just git + sqlite3 + bash. Zero API cost.
• Idempotent — safe to run init multiple times
• Non-destructive — merges into existing .claude/settings.json and CLAUDE.md without touching other content
• Incremental — only processes new commits after first run
• Branch-aware — context reflects your current branch
• Noise-filtered — lockfiles, generated files, and chore/ci/merge/build-bump commits are excluded automatically

What Gets Filtered

Commit noise (excluded from Recent Feature Commits):

• Merge commits (Merge branch ..., Merge pull request ...)
• Build bumps (chore: bump build, chore: release, chore: update deps)
• CI/style/format commits (ci:, style:)
• Package lockfile-only commits

File noise (excluded from Hot Files / Active Areas):

• package-lock.json, yarn.lock, Podfile.lock, *.lock
• node_modules/, dist/, build/, .expo/
• Root-level package.json, app.json

The Right Limit

The default --limit 15 means "show the 15 most recent meaningful commits" — not raw commits. If your project has 200 commits but only 15 are feat:/fix:/refactor:, you'll see those 15. Build bumps don't eat into your limit.

Why 15? Empirically derived from /dream consolidation runs: sessions reading 12–15 feature commits perform as well as sessions reading 260 raw commits, at a fraction of the token cost.

Requirements

• git (obviously)
• sqlite3 (available on macOS/Linux by default)
• bash

That's it. No Node, no Python, no npm install.

Philosophy

If you commit frequently with descriptive messages — and you should — your git history is already the best documentation of what happened and why. git-memory just makes that history legible to your AI coding assistant at the moment it matters most: session start.

The signal is in feat: and fix:. The noise is in chore: bump build 97. git-memory knows the difference.

Ma ka hana ka ʻike — In working, one learns. 🌺🤙🏼🚀

---

Author: @soderalohastrom License: MIT

Changelog

v1.1.0

• Recent Activity → Recent Feature Commits: only feat:, fix:, refactor:, perf:, add, implement, ship, revert patterns shown. Merge commits, chore: bump, ci:, style: silently excluded.
• Hot Files: lockfiles, node_modules/, dist/, build/, .expo/, and root package.json/app.json excluded from churn counts.
• Active Areas: same lockfile exclusions applied.
• --limit N now means "max meaningful commits shown" (default 15, was 20 raw commits).
• Motivated by /dream consolidation results: 12 feature commits > 260 raw commits for LLM context quality.

v1.0.0

• Initial release: SQLite-backed git history indexing, CLAUDE.md auto-update, SessionStart hook.