crabpath 3.0.0

Memory graph for AI agents that learns what to retrieve — and what to suppress.

🦀 CrabPath

Memory graph for AI agents. Learns what to retrieve — and what to suppress.

Why?

• RAG loads by similarity. CrabPath loads by learned routes.
• Positive outcomes strengthen paths. Negative outcomes create inhibitory edges.
• The graph gets smarter with every query. Unused paths decay.

Install

pip install crabpath

Python 3.10+. Zero dependencies.

Quick Start

from crabpath import Graph, Node, Edge, traverse, TraversalConfig
from crabpath.learn import apply_outcome

g = Graph()
g.add_node(Node("a", "Deploy to production"))
g.add_node(Node("b", "Rollback procedure"))
g.add_edge(Edge("a", "b", 0.7))

result = traverse(g, seeds=[("a", 1.0)], config=TraversalConfig(max_hops=3))
print(result.fired)

apply_outcome(g, result.fired, outcome=1.0)

For AI Agents (CLI)

• crabpath init --workspace <dir> --output <path>
• crabpath query <text> --graph <path> --index <path> [--top N] [--json]
• crabpath learn --graph <path> --outcome <float> --fired-ids <id1,id2,...>
• crabpath health --graph <path>

How It Works

• split_workspace scans markdown, creates nodes, and sibling edges.
• VectorIndex stores caller-provided embeddings and performs cosine search.
• traverse follows reflex/habitual/dormant tiers with edge damping.
• apply_outcome applies policy-style updates and Hebbian co-firing.
• apply_decay slowly forgets weak unused edges.
• autotune suggests safe defaults from graph health metrics.

Links

Paper / GitHub / PyPI