Stop losing context. Stop fighting compaction. Make Claude remember.
A lightweight system for managing Claude's memory across sessions without external dependencies.
Claude gets worse as conversations grow:
- Context bloat: 50K+ tokens of history slows everything down
- Compaction loss: Important decisions get summarized away
- Session amnesia: Each new session starts from scratch
- No handoff: Work gets repeated because Claude forgot what it learned
Three conventions that fix this:
| Pattern | What It Does |
|---|---|
| CLAUDE.md | Teaches Claude your project's capabilities and rules |
| Checkpoints | Saves session state to JSON, restores on start |
| Progressive Loading | Loads context in levels (100 → 500 → full tokens) |
No servers. No databases. Just files and conventions.
Click "Use this template" above to create your own repo, or clone directly:
git clone https://github.com/0xatem/context-kit.git
cd context-kitcp -r templates/ your-project/.claude/
cp templates/CLAUDE.md your-project/Edit CLAUDE.md in your project root with your project's details.
./scripts/checkpoint.sh init "Starting new feature"Tell Claude at session start:
"Read .claude/checkpoint.json and CLAUDE.md, then continue from where we left off."
context-kit/
├── README.md # You are here
├── SETUP.md # Detailed setup guide
├── docs/
│ ├── checkpoint-schema.md # JSON schema + guidelines
│ └── progressive-loading.md # Level 1/2/3 approach
├── scripts/
│ ├── checkpoint.sh # Checkpoint CRUD operations
│ └── session.sh # Session start/end helpers
└── templates/
├── CLAUDE.md # Project template
├── .claude/
│ └── checkpoint.json # Example checkpoint
└── workflows/
└── README.md # PR review, debugging, research patterns
Pre-built patterns for common tasks:
| Template | Use Case |
|---|---|
| PR Review | Multi-pass code review |
| Bug Investigation | Systematic debugging |
| Research Synthesis | Deep research with sources |
| Refactor Planning | Safe refactoring with checkpoints |
| Feature Planning | New feature design |
See templates/workflows/ for all templates.
Don't dump everything into context. Load in levels:
| Level | Tokens | Contains |
|---|---|---|
| 1 | ~100 | Project name, status, next task |
| 2 | ~500 | Summary, top 5 tasks, key decisions |
| 3 | Full | Everything (only for complex planning) |
Start at Level 1. Escalate only when needed.
Q: Do I need to run any servers? A: No. Everything is files. Read them, write them, done.
Q: Does this work with Claude.ai (web)? A: Yes. Copy checkpoint content into the conversation or use Projects.
Q: What about Claude Code's /compact command? A: Compaction summarizes aggressively. Checkpoints preserve what matters.
MIT - Use freely, modify freely, share freely.