Zaxy 0.4.0: Coordinator Memory For Agent Teams

Zaxy 0.4.0 release header

Zaxy 0.4.0 is the release where the project stops being only an agent memory system and becomes a coordinator memory layer for multi-agent work.

Spinning up agents is no longer the hard part. Worktrees, containers, task runners, and agent CLIs can already isolate ten workers. The hard part starts when those workers return: one found the current source of truth, one found an older rule, two duplicated each other, one made an unsupported claim, and the project now needs a coherent answer.

Zaxy 0.4.0 is built around that coordination problem.

The New Product Shape

Zaxy Coordinate introduces a parent mission plus isolated worker sessions:

parent mission
  -> worker sessions
  -> assignments
  -> cited findings
  -> conflict and stale-claim review
  -> accepted merge-back
  -> replayable project history

That is the important distinction. Worker sessions can investigate freely without turning every hypothesis into shared truth. The parent mission becomes the accepted state of the project. Rejected, stale, duplicated, and conflicting claims stay auditable, but they do not pollute the trusted checkout by default.

This changes Zaxy's core claim:

Isolated workers. Cited findings. Conflict detection. Accepted merge-back into one replayable project history.

That is a sharper product than generic persistent memory. It is designed for teams of agents that need a reliable control plane, not just a larger scratchpad.

What Ships In 0.4.0

Zaxy Coordinate

The release adds first-class coordination primitives for missions, workers, assignments, findings, decisions, promotions, handoffs, and approval packets. The Python and CLI surfaces expose a coordinator workflow that can start a mission, create scoped workers, collect evidence-backed findings, review those findings, and promote accepted state into the parent mission.

The key behavior is accepted-state checkout. A parent mission checkout should answer from promoted project state, while pending worker claims remain visibly separate. For agent teams, that boundary is the difference between useful parallelism and shared-memory contamination.

CoordinationBench

Zaxy 0.4.0 also introduces CoordinationBench as the benchmark lane for this new direction. CoordinationBench asks whether a system can turn multiple isolated worker outputs into one governed parent history.

The scorer measures accepted-finding precision and recall, conflict precision and recall, stale-claim rejection, duplicate consolidation, evidence coverage, parent-checkout answerability, citation coverage, Eventloom replayability, token estimates, and coordination latency.

The release deliberately keeps the benchmark story honest:

Lane	Score	What it means
CoordinationBench v1 public lane	`1.000`	First-party public-label reproducibility result.
CoordinationBench v1-scale public lane	`1.000`	First-party public-label scale reproducibility result.
Public-derived holdout mean	`0.606`	Current frozen-adapter generalization baseline.

The holdout result is the product signal to watch. It shows that the replay-backed coordination layer is strong at accepted-state precision and duplicate consolidation, while the frozen adapter still needs better source-aware final answering, stale-source interpretation, and broader conflict detection across public-derived cases.

Zaxy should not market a representative perfect CoordinationBench score. The public v1 scores prove the adapter can satisfy the public contract; the 0.606 holdout mean shows where the real next work is.

Embedded Graph Runtime

The plain local path now centers on the embedded Kuzu projection backend. Neo4j, pgGraph, and LatticeDB remain useful sidecar or evaluation tracks, but the release direction is zero-friction local memory: install Zaxy, initialize the workspace, and get Eventloom plus a rebuildable local graph projection without requiring an external graph database for ordinary use.

That matters for adoption. Agent memory only helps if it is present at the start of work. A sidecar-first setup creates friction exactly where users need the system to disappear.

The backend benchmark gates now protect that embedded default with reproducible reports, fingerprint checks, citation coverage, token-efficiency checks, and latency budgets. Optional backends remain explicit, not accidental.

Memory Bootstrap And Checkout

The model-facing memory contract is also tighter in 0.4.0. Zaxy exposes Memory Bootstrap and Memory Checkout as first-class ways to tell an agent what memory exists, what can be trusted, what needs refresh, and how to cite or reinforce retrieved context.

The checkout object now carries diagnostics for answerability, confidence, source lanes, citation coverage, unsupported compacted context, warnings, and feedback guidance. That makes memory less magical and more operational: the agent can see whether it is answering from current cited state, stale state, or insufficient memory.

Deterministic Capture And Activation

Zaxy 0.4.0 continues the local-first capture work. Deterministic capture records session turns, tool calls, commands, and file edits into Eventloom so the memory system can reconstruct what happened instead of relying only on manually written summaries.

The release also treats activation as a product KPI. Capturing work after the fact is not enough; useful memory needs to be checked out before substantive work starts. The doctor, hook-status, and capture-soak paths now make that visible with freshness, coverage, and token-efficiency diagnostics.

Why This Release Matters

Most agent memory products are converging on retrieval: store facts, search facts, inject facts. That is necessary, but it does not solve the coordination layer for agent teams.

Multi-agent work needs three extra properties:

Isolation, so worker hypotheses do not become global truth immediately.
Governance, so findings are accepted, rejected, or deferred with evidence.
Replay, so the final project state can be audited back to the events that produced it.

Zaxy already had the substrate: Eventloom append-only history, temporal graph projection, citations, Memory Checkout, MCP, and local capture. Version 0.4.0 connects those pieces into a product direction that is easier to explain and harder for flat memory systems to copy.

The release is not claiming that Zaxy has solved all of coordination. It is claiming the right architecture for the problem: parent mission state, worker session isolation, evidence-backed promotion, and a benchmark that measures the actual failure modes.

Install Or Upgrade

The PyPI distribution is zaxy-memory; the import package and console command remain zaxy.

pip install -U zaxy-memory
zaxy init
zaxy doctor

For local Codex onboarding:

zaxy init --preset local-codex

For the Coordinate docs:

zaxy coordinate --help