Changelog¶

All notable changes to this project are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning.

[Unreleased]¶

House-rules benchmark (benchmarks/house_rules/): an agent learns team conventions from verifier feedback alone. Held-out pass rate went from 0% to 91% after 24 runs. Includes a chart generator and the run history.
examples/demo_learning.py: watch an agent fail, learn and pass in about a minute.

Better learning quality, found with the benchmark:
The curator and evolver never state specific values (limits, names, numbers) that the evidence doesn't contain.
They keep required formats (markers, signatures, prefixes, limits) verbatim, and use placeholders only for run-specific values.
Rules describe what the output must contain, not research steps the agent may be unable to perform.
New evidence that contradicts a rule updates that rule instead of adding a second one.

All three registries now publish without stored credentials: PyPI and npm through trusted publishing (OIDC), GHCR through the workflow token. Every artifact carries signed provenance.

The OpenClaw plugin ships compiled JavaScript (dist/), so openclaw plugins install @myelinate/openclaw-plugin works from npm. 0.2.0 contained only TypeScript sources, which OpenClaw loads only from local checkouts.

Install from the registries: pip install myelinate, ghcr.io/notaaaronfr/myelinate, and @myelinate/openclaw-plugin on npm.

Documentation site (MkDocs Material, published to GitHub Pages): quickstart, concepts, integration guides, Docker and Kubernetes deployment, a production checklist, operations guides (security and SSO, governance, observability, data and privacy, scaling and HA, backup and upgrades, cost control), troubleshooting, FAQ, and configuration and API references generated from the code.
Kubernetes manifests (deploy/kubernetes/), verified on k3s.
Runnable examples (examples/).
myelinate healthcheck for container probes. Standalone workers now get a heartbeat that is independent of job length, graceful SIGTERM shutdown, JSON logs, and Prometheus metrics on MYELINATE_WORKER_METRICS_PORT (9464).
GET /v1/episodes/latest, so stateless clients can attach feedback to a session's last turn.
New settings: MYELINATE_AUDIT_RETENTION_DAYS, MYELINATE_DB_POOL_SIZE, MYELINATE_DB_MAX_OVERFLOW and MYELINATE_WORKER_METRICS_PORT.
OIDC scopes are now also read from Auth0 permissions and Keycloak realm_access.roles.

agent_id="*" (knowledge shared tenant-wide) can only be written through review-scoped skill endpoints.
OIDC outside dev requires both MYELINATE_OIDC_ISSUER and MYELINATE_OIDC_AUDIENCE.
myelinate purge also removes old recall logs, finished jobs and (per MYELINATE_AUDIT_RETENTION_DAYS) audit entries.
Redaction placeholders are now [REDACTED <label>].
The insights field learning_cost_30d is renamed tenant_learning_cost_30d.
Docker Compose services restart automatically. The image healthcheck works for both gateways and workers.

Recall injects a skill's rules in their authored order.
Gateways and workers starting together on an empty PostgreSQL no longer race on schema creation.
When the learning LLM is unavailable or the budget is exhausted, the API returns 503 or 429 with Retry-After, instead of 500.
A missing Anthropic API key fails individual learning steps instead of whole jobs.
Memories that can't be reconciled without the LLM are held for review instead of being dropped.
Redaction no longer mangles secret labels or swallows text after card numbers.
Rate-limit responses include Retry-After.
Changeset and experiment metrics count transitions from canary promotion, queue release and rollback.
The Docker image runs standalone with SQLite (data on a /data volume).

Gateway with REST, MCP and OTLP ingestion; adapters for Microsoft Agent Framework, OpenClaw and any framework.
Skills as individually credited rules, evidence-gated reflection and curation, and GEPA-style evolution.
Bayesian canary rollout and a holdout group measuring causal lift.
Memory with reconciliation and validity intervals, and per-user profile cards.
Multi-tenancy, scoped API keys, OIDC, audit log, PII redaction, erasure, budgets and idempotent ingestion.