4.3 KiB
███╗ ███╗███████╗███╗ ███╗ █████╗ ██████╗ ██████╗ █████╗
████╗ ████║██╔════╝████╗ ████║██╔══██╗██╔══██╗██╔══██╗██╔══██╗
██╔████╔██║█████╗ ██╔████╔██║███████║██████╔╝██████╔╝███████║
██║╚██╔╝██║██╔══╝ ██║╚██╔╝██║██╔══██║██╔══██╗██╔══██╗██╔══██║
██║ ╚═╝ ██║███████╗██║ ╚═╝ ██║██║ ██║██████╔╝██║ ██║██║ ██║
╚═╝ ╚═╝╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝╚═════╝ ╚═╝ ╚═╝╚═╝ ╚═╝
Intuition-driven control plane for agent memory & action selection.
🧠 What is memabra?
Most agents memorize. memabra intuits.
memabra is a local-first, observable, trainable, and replayable control plane for agent memory and action orchestration.
Instead of acting like a dusty filing cabinet, memabra functions as a meta-cognitive controller: given any task, it rapidly decides whether to answer directly, recall memory, load a skill, or invoke a tool — then learns from outcomes to sharpen those instincts over time.
- 🏠 Local-first — no cloud lock-in, your data stays on disk
- 📊 Observable — every decision is tracked, versioned, and inspectable
- 🎓 Trainable — online learning loop improves routing automatically
- 🔄 Replayable — replay trajectories, audit decisions, roll back versions
⚡ Quick Start
git clone https://github.com/TacitLab/memabra.git
cd memabra
python -m venv venv
source venv/bin/activate
pip install -e ".[dev]"
1. Peek under the hood
memabra --help
2. Run a safe dry-run evaluation
See the full workflow without actually promoting a new router version:
memabra run --dry-run --format text
3. Check system pulse
memabra status --format text
4. Inspect your router lineage
memabra version list --format text
5. Time-travel (rollback)
memabra version rollback <version-id> --format text
🎮 CLI Commands
| Command | Description |
|---|---|
memabra run |
🚀 Execute the online learning workflow |
memabra status |
💓 Show current system health & metrics |
memabra version list |
📜 List all saved router versions |
memabra version rollback <id> |
⏪ Roll back to a specific version |
🖨️ Operator-Friendly Output
By default, memabra speaks JSON. For humans, add --format text:
memabra run --dry-run --format text
Sample output:
Memabra online learning result
Summary
Report ID: report-58f9f22
Skipped: no
Promoted: yes
Dry run: yes
Baseline
Reward: 0.7200
Error rate: 0.1200
Latency (ms): 145.0000
Challenger
Reward: 0.8100
Error rate: 0.0800
Latency (ms): 132.5000
Deltas
Reward delta: 0.0900
Error rate delta: -0.0400
Latency delta (ms): -12.5000
Decision
Accepted: yes
Reason: challenger improved reward and reduced error rate
✅ Normalized booleans (yes/no/none)
✅ Fixed-precision metrics for easy comparison
✅ Sectioned layout — Summary → Baseline → Challenger → Deltas → Decision
🧪 Running Tests
pytest tests/ -q
Current status: 126 passed 🟢
📚 Documentation
docs/PROGRESS.md— Capability roadmap & what's shippeddocs/DEMO.md— Hands-on walkthroughs & examplesdocs/ARCHITECTURE.md— System design & mental model
🏷️ License
MIT — use it, break it, improve it.
Built with caffeine and curiosity. ☕