crabpath 5.0.1

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

CrabPath

CrabPath is a pure in-memory graph engine for retrieval routing that can learn from feedback and prioritize what to execute next, while staying independent of any model provider.

Install

pip install crabpath

Python API (pure callbacks)

from crabpath import Graph, Node, Edge, VectorIndex, split_workspace, traverse, apply_outcome

# 1) Build graph + texts from workspace
graph, texts = split_workspace("./workspace")

# 2) Caller owns embeddings
embed = lambda text: [1.0, 2.0]  # your vector model, local or remote
index = VectorIndex()
for node_id, text in texts.items():
    index.upsert(node_id, embed(text))

# 3) Caller owns LLM callbacks (optional)
def route_fn(query: str, candidate_ids: list[str]) -> list[str]:
    return candidate_ids[:3]

def score_fn(system_prompt: str, user_prompt: str) -> str:
    return '{"scores": {"node-id": 1.0}}'

# 4) Query and learn
seeds = index.search([0.1, 0.2], top_k=8)
result = traverse(graph=graph, seeds=seeds, route_fn=route_fn)
apply_outcome(graph=graph, fired_nodes=result.fired, outcome=1.0)

CLI (no providers, pure stdin/stdout)

# Build graph, texts, and optional index from a callback
crabpath init --workspace ./workspace --output ./crabpath-data --embed-command 'python3 embed_cb.py'

# Build index only
crabpath embed --texts ./crabpath-data/texts.json --output ./crabpath-data/index.json --command 'python3 embed_cb.py'

# Query by keyword
crabpath query "how do i deploy" --graph ./crabpath-data/graph.json --top 5

# Query by vector payload file
crabpath query "noop" --graph ./crabpath-data/graph.json --index ./crabpath-data/index.json --query-vector-stdin < vec.json

# Optional route callback and query scoring callback wiring
cat /tmp/query.vec | crabpath query "deploy" --graph ./crabpath-data/graph.json --index ./crabpath-data/index.json --route-command 'python3 route_cb.py' --embed-command 'python3 embed_cb.py'

Batch callbacks

Batching is available for both embedding and callback APIs using the same CLI entry point:

crabpath init --workspace ./workspace --output ./crabpath-data --embed-command 'python3 embed_batch.py'
crabpath query "deploy" --graph ./crabpath-data/graph.json --route-command 'python3 route_batch.py' --json

Internally, ThreadPoolExecutor is used to parallelize single-item fallback while preserving batch callbacks when provided.

Other pure graph commands

crabpath learn --graph graph.json --outcome 1.0 --fired-ids a,b,c
crabpath replay --graph graph.json --sessions ./sessions/*.jsonl
crabpath health --graph graph.json --json
crabpath merge --graph graph.json --json
crabpath connect --graph graph.json --json
crabpath journal --stats

What this project is

CrabPath is a library, not a hosted service. It does no network calls and persists no credentials. Callers pass in embedding and routing callbacks; CrabPath only manages graph logic, scoring plumbing, and state updates.

Paper

Technical details: https://jonathangu.com/crabpath/