Crewbrain
a shared memory graph for agent teams
decisions with reasons, citations, human approval
Graph-only answers - no paid model. Built on FalkorDB. Demo data is the fictional Harborlight Outfitters company; nothing here is a real customer.
What exists today
Shipped
Decision graph + as-of time travel
Decisions are nodes linked to chosen and rejected options, reasons, supporting passages, sources and deciders. Facts are superseded, never overwritten, so "what did we believe on 1 Sep?" still works.
OpenShipped
Cited Ask with honest abstention
Ask traces a subgraph and answers only from it, citing passage ids. A validator rejects any citation outside the traced subgraph; with no recorded decision it says "no recorded decision" instead of guessing.
OpenShipped
Document ingest (text PDFs included)
Pasted or uploaded .txt/.md/.json/.csv/.log and text PDFs chunk into Source and Passage nodes with hashed idempotency, then open reviewable proposals. Scanned/image-only PDFs have no OCR and are rejected honestly. DOCX is not supported yet.
OpenShipped
Requirement-first Start (unified run)
Start runs the real pipeline from a plain-language requirement: Ask first, then task creation and a real agent claim, with proposals in-run and a re-ask after approval. It stops honestly where human approval or manual work is required.
OpenShipped
Agent task claiming (no double-claims)
One atomic compare-and-set Cypher query claims a task; at most one concurrent agent wins. Measured: 8 agents x 50 tasks, 50 claimed, 0 double claims.
OpenShipped
Per-tenant sandboxes
Each tenant is its own FalkorDB graph (t_acme, t_sbx_*). Visitor sandboxes live 24 h, are capped at 20, and show a GRAPH.MEMORY USAGE meter per graph.
OpenShipped
Handoffs with evidence + context cards
A handoff without at least one existing Memory, Passage or Decision is rejected. The receiver's next session starts from a context card built from the graph.
OpenHow it works
- 1.Ingest a document or paste a requirement - Ask memory first.
- 2.Trace the graph and propose changes (proposals only, never auto-applied).
- 3.A human approves or rejects each proposal (agents cannot approve).
- 4.Agents claim tasks atomically and hand off with evidence; re-ask the original question.
Honest limits
- Retrieval today is alias and entity matching with fuzzy Jaro-Winkler for decisions, plus plain keyword matching over passage text for source recall - not semantic search. The declared vector and fulltext indexes exist but are unused.
- Agents claim tasks but do not execute beyond claiming. Execution and approvals stay human.
Demo
Watch the demo
The demo video link is in the README (demo/crewbrain-demo.mp4). It is not served from this site.