Zaxy Coordinate - coordinator memory for agent teams
Coordinator memory for agent teams
Run many agents without losing the plot.
Zaxy Coordinate gives multi-agent projects a parent mission, isolated<br>worker sessions, cited findings, conflict diagnostics, approval packets,<br>and accepted merge-back into one replayable project history.
Install locally<br>Read Coordinate docs<br>View CoordinationBench
1797 tests<br>91.96% coverage<br>ruff clean<br>mypy clean<br>PyPI 0.4.0
The wedge
Worker-local claims are not project truth.
Spawning agents is easy. The hard part is turning ten isolated<br>investigations into one trustworthy state of work. Zaxy keeps worker<br>sessions separate, records findings with evidence, marks stale and<br>conflicting claims, and promotes only accepted findings into the parent<br>mission session.
Parent mission<br>The coordinator owns the accepted project history, decisions, handoff, and Memory Checkout state.
Worker sessions<br>Agents investigate in isolated Eventloom logs, so exploration does not contaminate authoritative memory.
Approval packets<br>Human or coordinator-agent review accepts, rejects, defers, or promotes findings with cited provenance.
Architecture
Event-sourced coordination, not a shared scratchpad.
Mission objective and parent state
Workers isolated sessions
Findings evidence and confidence
Review conflicts, stale claims, approvals
Checkout accepted cited context
Eventloom remains the append-only source of truth. The embedded Kuzu<br>projection is the default local graph runtime; Neo4j, pgGraph, LatticeDB,<br>and Pathlight are advanced integration tracks. MCP exposes both memory<br>primitives and first-class coordination tools.
Interfaces
CLI, MCP, dashboard, and adapters all speak the same coordination model.
zaxy coordinate briefReplay mission state, accepted findings, conflicts, stale claims, missing evidence, and next decisions.
coordination_checkoutReturn accepted parent state plus diagnostic worker-local findings without treating them as truth.
coordination_approval_packetGenerate a reviewable payload for accept/reject/defer/promote decisions.
memory_checkoutKeep the lower-level model-facing contract with answerability, required_action, current_citation_count, and memory_feedback guidance.
CoordinationAdapterDependency-light Python wrapper with LangGraph and CrewAI helper paths.
dashboard --enable-coordinate-reviewOpt-in human review controls on top of replay-backed mission state; read-only remains the default.
Benchmark evidence
CoordinationBench is the coordination benchmark.
Zaxy Coordinate now has a frozen first-party adapter for the external<br>CoordinationBench scorer. CoordinationBench v1 and v1-scale score<br>accepted-finding precision and recall, stale rejection, duplicate<br>consolidation, evidence grounding, and answerability. Zaxy scores<br>1.000 on those public reproducibility lanes, then lands at a 0.606<br>mean on public-derived holdout packs. Treat the holdout number as<br>the real current baseline.
Holdout mean<br>0.606 overall
Frozen-adapter public-derived packs expose the next work: final answering, stale interpretation, and broader conflict detection.
Public lanes<br>1.000 overall
The v1 and v1-scale scores are first-party public-label reproducibility results, not representative leaderboard claims.
Competitors<br>disclosure only
Mem0, Agent Memory, and ActiveGraph adapters are listed as not_run until pinned same-harness runner manifests exist.
CoordinationBench Results
Frozen adapter result: CoordinationBench v1 shows 1.000 accepted precision versus 0.200 flat precision, while public-derived holdouts show the current generalization gap.
Lane<br>overall<br>accepted precision<br>conflict recall<br>stale rejection<br>answerability
v1-audited<br>1.000<br>1.000<br>1.000<br>1.000<br>1.000
v1-scale<br>1.000<br>1.000<br>1.000<br>1.000<br>1.000
public-derived holdout mean<br>0.606<br>0.846<br>0.375<br>0.000<br>0.000
source-weight baseline v1<br>0.411<br>0.683<br>0.800<br>0.333<br>0.600
Markdown notes<br>0.400<br>0.000<br>0.000<br>0.000<br>0.000
BM25 worker logs<br>0.333<br>0.000<br>0.000<br>0.000<br>0.000
Flat transcript<br>0.200<br>0.000<br>0.000<br>0.000<br>0.000
CoordinationBench reportFrozen adapter and holdout results<br>Benchmark hubLongMemEval, backend shootout, disclosures<br>Coordinate roadmapProduct direction and adapter contract<br>LongMemEval guardrailSupporting memory retrieval evidence
Install
Five-minute local smoke test, then expose coordination through MCP.
pipx install zaxy-memory<br>zaxy init<br>zaxy memory log --eventloom-path .eventloom --limit 5<br>zaxy memory bootstrap --eventloom-path .eventloom<br>zaxy doctor --eventloom-path .eventloom<br>zaxy coordinate start "ship auth refactor" --mission auth-main<br>zaxy coordinate worker create --mission auth-main --worker auth-api<br>zaxy coordinate assign --mission auth-main --worker auth-api "trace failures"<br>zaxy coordinate brief --mission auth-main<br>zaxy coordinate checkout --mission auth-main
What happens when you run init<br>Zaxy writes `.env.local`, records session genesis and heartbeat, checks graph posture, and prints the MCP command or config path.
What...