crabpath 2.3.2

🦀 Memory store for AI agents that learns what to retrieve — and what to suppress — from experience.

🦀 CrabPath

CrabPath is a memory store for AI agents that learns what to retrieve — and what to suppress — from experience.

Why?

• Static context loading wastes tokens because you load too much every turn.
• Classic RAG can’t learn from feedback, so retrieval stays similarity-only.
• CrabPath tracks which retrieval paths worked, and builds a graph of learned routes.

Install

python3 -m venv ~/.crabpath-env && source ~/.crabpath-env/bin/activate
pip install crabpath            # PyPI
clawhub install crabpath        # or ClawHub (OpenClaw agents)

# For embeddings (strongly recommended):
pip install crabpath[openai]     # or: pip install crabpath[google]

Zero required dependencies. Python 3.10+. macOS Homebrew Python needs a venv (PEP 668).

LLM Access (recommended)

CrabPath's smart routing needs a real LLM. Without one, it falls back to basic retrieval (like RAG). Provide access via any of:

# Option 1: Direct API key (easiest)
export OPENAI_API_KEY=sk-...          # also checks ~/.env automatically

# Option 2: Any OpenAI-compatible endpoint (most flexible)
export CRABPATH_LLM_URL=http://localhost:8080/v1/chat/completions
export CRABPATH_EMBEDDINGS_URL=http://localhost:8080/v1/embeddings
export CRABPATH_LLM_TOKEN=your-token  # if endpoint needs auth

# Option 3: Local Ollama (free, private)
ollama pull nomic-embed-text          # embeddings
ollama pull llama3                     # routing

Works with: OpenAI, Gemini, Ollama, LiteLLM, vLLM, OpenRouter — anything OpenAI-compatible.

Without LLM access, CrabPath still works using local TF-IDF embeddings and heuristic routing — but you're missing the learned routing that makes it better than RAG.

Quick Start (60 seconds)

from crabpath import Node, Edge, Graph, activate, learn

g = Graph()
g.add_node(Node("timeout", "Deployment timed out"))
g.add_node(Node("rollback", "Rollback and restore"))
g.add_node(Node("debug", "Inspect logs"))
g.add_edge(Edge("timeout", "rollback", 0.6))
g.add_edge(Edge("timeout", "debug", 0.4))

result = activate(g, seeds={"timeout": 1.0})
learn(g, result, outcome=1.0)  # reinforces paths that fired
print([node.id for node, energy in result.fired])

For AI Agents (3 commands)

crabpath init --workspace ~/.openclaw/workspace --sessions ~/.openclaw/agents/main/sessions/
crabpath install-hook --agent-workspace ~/.openclaw/workspace
crabpath query 'how do I deploy' --graph ~/.crabpath/graph.json --index ~/.crabpath/embed.json --top 8 --json

Which Interface?

Interface	Status / Use
CLI (agents)	Primary agent-facing interface; JSON I/O for shell workflows.
MemoryController (Python)	Recommended direct integration for Python apps.
Adapter	Deprecated legacy bridge; prefer CLI or MemoryController.

How It Works

• Documents are split into nodes and edges become weighted pointers.
• Reflex edges (>0.8) auto-follow with near-zero overhead.
• Habitual edges (0.3-0.8) go through normal routing policy.
• Dormant edges (<0.3) are suppressed by default.
• Positive outcomes (+1) strengthen paths; negative outcomes (-1) create inhibitory edges.
• Decay drops unused connections, while the autotuner keeps graph routing healthy.

Key Results

Metric	Result
Context reduction	90-99%
Negation accuracy	1.0 vs 0.0 (BM25)
Internal tests	360
Required deps	Zero

Full benchmark details: docs/research/

When NOT to Use CrabPath

• Simple static-document RAG without feedback loops (use a vector DB).
• Very small codebases (< 10 files) with insufficient structure to learn recurring routes.
• One-off questions with no recurring retrieval patterns.

Links

• Paper: jonathangu.com/crabpath/
• ClawHub: clawhub.ai
• PyPI: pypi.org/project/crabpath/

License

Apache 2.0