Production memory for agent teams that need receipts.
Zaxy turns agent context into an auditable project memory fabric. It captures
parent missions, worker sessions, tool observations, cited findings, conflict
review, approval packets, and accepted merge-back into one durable history that
can be queried, replayed, and inspected.
Under the hood, Zaxy uses Eventloom append-only JSONL as the source of truth and
an embedded LadybugDB graph projection for local reasoning. It is built for agents
that need to remember what happened, cite where it came from, and avoid turning
project state into a pile of markdown files and vector chunks.
Event-sourced, auditable agent memory over MCP: cited checkout, replayable provenance.
2026-11-10
ดาว
—
เครื่องมือ
—
อัปไทม์
—
The embedded LadybugDB graph projection is the default local runtime.
The plain install uses embedded LadybugDB. Install zaxy-memory[neo4j] only for the
optional Neo4j sidecar, and zaxy-memory[pathlight] only for Pathlight tracing.
Why It Matters
Auditable memory: every accepted fact can point back to Eventloom history.
Agent-team coordination: parent and worker sessions stay isolated until
findings are reviewed and merged.
Local-first runtime: the default path uses embedded LadybugDB, no Neo4j sidecar.
MCP-native integration: Codex, Claude Code, Cursor, VS Code, Hermes Agent,
LangGraph, CrewAI, and AutoGen can use the same memory interface.
Benchmark claims withdrawn (2026-07-03). The prior LongMemEval numbers
were withdrawn: they were produced in oracle mode (mean ~1.9 candidate
sessions per question, so Recall@5/citation-coverage were ~1.0 by
construction, not by retrieval) and the preference-question scores rested on a
hardcoded answer table that has since been removed. Zaxy does not currently
publish a LongMemEval score. A real, full-haystack LongMemEval run is planned;
until it lands, treat the earlier 0.956/0.910/1.000 figures as retracted.
The Harvey LAB claim is pending the same audit.
The PyPI distribution is zaxy-memory; the import package and console command
are still zaxy. Bare zaxy init sets up the local embedded graph posture,
repo-local profile, deterministic capture config, genesis event, heartbeat, and
MCP guidance. For Codex, the printed activation launcher starts the managed
capture watcher when the local capture config is present; pass --capture start
only when you want init itself to start the watcher before opening Codex. The
default human output is compact and action-first; add --verbose when you need
the full setup diagnostics, optional checks, fallback commands, resume guidance,
and notes.
For automation, zaxy init --json keeps the raw onboarding fields and adds
setup.status, setup.issues, setup.pending, readiness.status,
readiness.reasons, readiness.actions, and structured
readiness.action_items for both commands and non-command review tasks. Each
structured action carries label, command, original source, and hints
for compact-output tips such as activation <task> replacement and path-stable
command guidance. Installers can render those tips without parsing prose. It also
includes setup.summary, readiness.summary,
readiness.required_action_count, and readiness.reason_count, so client UIs
can render compact status without parsing human output. It also
separates readiness.blocking_diagnostics from
readiness.non_blocking_diagnostics so scripts can distinguish setup
completion, required actions, and advisory doctor warnings before relying on
live memory.
For Codex, zaxy init --codex-mcp-install auto is the default. It writes or
reuses the user-level Codex MCP config when that can be done without replacing
an existing zaxy server entry. If no safe config target exists, it prints the
copyable codex mcp add command. If an existing zaxy entry differs, it asks
you to review that config before replacing it because Codex can silently replace
servers with the same name. Use an explicit mode when you need to force one side
of that decision after review:
zaxy init --codex-mcp-install user
# or: zaxy init --codex-mcp-install command
Both Codex paths keep the server workspace-neutral. After init, start or
restart Codex through the printed zaxy activate codex ... --launch command so
the MCP server list and Zaxy activation packet are loaded together. The printed
command includes explicit --eventloom-path and --workspace-root values, so
it still targets the initialized repo when copied from another shell.
Run the single-agent memory example:
python examples/single_agent_memory.py
Your local data lives under .eventloom/ as one append-only JSONL file per
session.
For repository development, use pip install -e ".[dev]", ./scripts/setup.sh,
and zaxy status. Start Docker sidecars only for integration tests or explicit
backend comparisons. Production setup writes Docker secret files under
./secrets/; see docs/deployment.md.
Zaxy also includes an observe-only OpenAI-compatible packet analyzer for model
call provenance. It forwards packets to one configured upstream endpoint and
records llm.packet.completed events to Eventloom without acting as a router.
See LLM Packet Analyzer.
Public Site and Documentation
Public static site: site/index.html
Why Zaxy: docs/why-zaxy.md
Getting started: docs/getting-started.md
MCP quickstart: docs/mcp-quickstart.md
Architecture: docs/architecture.md
Configuration: docs/configuration.md
MCP interface: docs/mcp.md
Memory export contract: docs/export-contract.md
Eventloom contract: docs/eventloom.md
Graph schema: docs/graph-schema.md
Retrieval: docs/retrieval.md
Benchmarks: docs/benchmarks.md
LLM packet analyzer: docs/packet-analyzer.md
Embeddings: docs/embeddings.md
Security: docs/security.md
Operations and deployment: docs/operations.md, docs/deployment.md, docs/runbook.md
Session sharding: One Eventloom log per agent/session, with a shared graph.
MCP-native: Drop-in memory for any MCP-compatible agent framework over stdio or SSE.
Observable: Optional Pathlight traces, breakpoints, and diff support via zaxy-memory[pathlight].
Hardened local defaults: bounded MCP inputs, safe session IDs, no-sidecar embedded graph projection, and optional admin token support for replay/invalidation.
Project Structure
File
Purpose
src/zaxy/event.py
Eventloom JSONL I/O + hash chain integrity
src/zaxy/extract.py
Hybrid extraction engine + rule registry
src/zaxy/embedded_graph_store.py
Embedded LadybugDB projection store
src/zaxy/graph.py
Optional Neo4j bi-temporal wrapper via zaxy-memory[neo4j]
src/zaxy/query.py
Hybrid retrieval router
src/zaxy/mcp_server.py
MCP stdio/SSE server
src/zaxy/trace.py
Optional Pathlight observability hooks
src/zaxy/core.py
MemoryFabric orchestrator
src/zaxy/session.py
Per-session Eventloom log manager
src/zaxy/security.py
Shared validation and input bounds
src/zaxy/__main__.py
CLI (zaxy serve, zaxy replay, etc.)
Production Secrets
Zaxy supports Docker/Kubernetes-style secret files for sensitive settings:
Variable
Secret-file variant
NEO4J_PASSWORD
NEO4J_PASSWORD_FILE
MCP_ADMIN_TOKEN
MCP_ADMIN_TOKEN_FILE
PATHLIGHT_ACCESS_TOKEN
PATHLIGHT_ACCESS_TOKEN_FILE
Direct environment variables take precedence over their *_FILE variants.
Use docker-compose.prod.yml as the production compose baseline.
Development
Tests first (Karpathy rule). Every public function has a test.
Unit tests mock external services. Integration tests use Docker for optional sidecar backends.
Coverage gate: ≥92% enforced by CI.
Lint/format: ruff. Types: mypy.
# Run full suite with coverage gate
pytest
# Run integration tests (requires Docker)
./scripts/generate-certs.sh .certs
docker compose --profile integration up -d neo4j-test neo4j-tls
pytest -m integration --no-cov
# Lint and type-check
ruff check src tests
mypy src
# LongMemEval benchmark numbers are WITHDRAWN (see the note at the top of this
# README). The `--dataset .cache/.../longmemeval_oracle.json` path only exercises
# the answer step over pre-selected gold sessions (oracle mode) and does NOT
# measure retrieval on the full LongMemEval haystack; do not publish figures from
# it as a LongMemEval score. A real full-haystack harness is TBD.
# Harvey LAB external memory-ablation comparison
# Consumes externally generated Harvey normalized-result artifacts for Zaxy;
# does not reuse LongMemEval statistics as legal-agent benchmark evidence.
# Current full external Harvey LAB evidence:
# reports/benchmarks/harvey-lab-memory-ablation/publishable-statistics.md
# reports/benchmarks/harvey-lab-memory-ablation/harvey-lab-benchmark.json
# 10/10 tasks, mean criterion pass rate 0.788, +0.184 vs regular/no-memory,
# +0.081 vs article-best task rows, 9/10 wins vs article-best rows.
# Production deployment preflight
scripts/validate-deployment.sh --root .
# Build and validate Python release artifacts
scripts/build-dist.sh --root .
# Verify local release metadata, PyPI Trusted Publishing, and LangGraph smoke
zaxy doctor --release-smoke
# Validate public site and documentation links
scripts/validate-docs.sh --root .
# Clean-repo beta UAT: install into a throwaway workspace and verify init,
# bootstrap, deterministic capture, doctor, and memory checkout.
scripts/beta-uat.sh
# Summarize beta readiness gates without external services.
zaxy doctor --beta-readiness
# Go-live release gate
scripts/release-check.sh --root .
The full suite must stay at or above 92% coverage before a sprint is complete.
Release Publishing
The PyPI distribution name is zaxy-memory because zaxy is already occupied
on PyPI. Published releases build from GitHub Actions and upload to
https://pypi.org/project/zaxy-memory/ using PyPI Trusted Publishing with
GitHub OIDC. The import package and console command remain zaxy.
Before publishing, run zaxy doctor --release-smoke to verify the package
version, changelog entry, release workflow, tokenless publishing posture, and
dependency-light LangGraph example.