0.2.0-harness-preview
OpenClaw-compatible
Markdown-native
Human-in-the-loop
GARL-ready

MemexLab for OpenClaw agents

OpenClaw gives the agent hands. MemexLab gives it a governed long-term mind — a markdown-native vault the agent operates through explicit skills, under schemas, evals, and a public/private boundary.

Agent quick-start Read the docs

# skills the agent invokes over the vault
memex-ingest    <source>      → sources/   (immutable, with provenance)
memex-markdown  <item>        → items/     (atomic, linked, cited)
memex-query     "<question>"  → cited answer
memex-brief     "<topic>"     → brief / memo
memex-evaluate                → quality report (retrieval · citations · dedup)
# writes are dry-run; a human approves; the private vault is never published

The split

Runtime vs. memory

OpenClaw provides the agent runtime and assistant surface — tools, context, and the ability to act. MemexLab provides the durable knowledge layer those actions read from and write back to: a plain-markdown vault that can be inspected, cited, evaluated, and governed. The agent doesn't accumulate memory in a chat log or an opaque index — it maintains files you own.

This is harness engineering applied to memory: reliability comes from the scaffolding — schemas, skills, evals, verification, governance — not the model. The model is swappable (Anthropic, OpenAI, …); the harness is what compounds.

Preview OpenClaw skill packaging is experimental in 0.2.0-harness-preview. The skills below are described at the product level; concrete packaging and runtime bindings are being validated and may change before a stable release.

The skills

Five skills the agent operates

Each skill is a named, reviewable operation over the vault. It reads from immutable sources and the markdown layer, writes a schema-shaped artifact, and leaves a trail.

memex-ingest

Brings a source into the vault as a traceable source note.

Writes sources/ — full or lightly-cleaned input.

Why every later claim can be traced to where it came from.

memex-markdown

Maintains the canonical markdown layer — atomic items with valid frontmatter and links.

Writes items/ — one thesis or idea per note.

Why knowledge stays human-readable, diffable, and re-linkable.

memex-query

Answers a question by retrieving over the vault and citing what it used.

Returns a cited answer (source paths + entity slugs).

Why answers are auditable and can be filed back to compound.

memex-brief

Synthesizes a brief or memo on a topic from the canonical layer.

Returns a structured brief that links to its sources.

Why a position is assembled from evidence, not re-derived.

memex-evaluate

Runs the eval query-set and rubric over the vault.

Returns a quality report — retrieval, citations, dedup, contradictions.

Why quality is measured over time, not assumed.

schemas & evals

Entity schemas define valid artifacts; eval sets + a rubric define "good."

Where in the public memexlab-engine repo (e.g. schemas/entity.schema.json, evals/).

Why the agent's output is validatable, not vibes.

Agent quick-start

Wire the skills into your agent

Preview The MemexLab skills, schemas, evals, and governance live in the public memexlab-engine repo. OpenClaw is the full agent surface (below); a minimal self-hosted runner is included if you'd rather run locally with no extra runtime. The memex CLI commands shown in the docs remain conceptual until that package ships.

Install OpenClaw

MemexLab's skills run inside an OpenClaw agent — a personal AI assistant runtime (MIT-licensed; Node 24 recommended, 22.19+ supported). Install it from npm and onboard:
```
npm install -g openclaw@latest   # or: pnpm add -g openclaw@latest
openclaw onboard --install-daemon
```
We link to OpenClaw rather than bundling it — see the official repo for current install steps and version support.
Add the skills

Place the MemexLab skills in your agent's skills directory so the OpenClaw runtime can discover memex-ingest, memex-markdown, memex-query, memex-brief, and memex-evaluate.
Point at a vault

Set the vault path (e.g. ~/Documents/Obsidian/<your-vault>/). Keep your real private vault separate from the public framework — the boundary is enforced by governance.
Operate

The agent runs skills: memex-ingest a source, memex-query a question, memex-brief a topic. Retrieval is deterministic; answers cite their sources.
Review before write

Writes are dry-run by default. A human reviews the proposed changes and approves — nothing is committed or published unsupervised.
Validate & compound

Run memex-evaluate to check quality, then file useful answers back into the vault so knowledge compounds instead of being re-derived.

No install Prefer no extra runtime? The self-hosted runner is a dependency-light Python loop that runs the same skills against your vault, with a switchable local-or-hosted backend — see the self-hosting guide.

Governance for agents

What an agent may read, write, and publish

Agent actions are bounded by a declarative policy (governance.yml, defined in the memex engine repo), so an autonomous tool can't quietly leak or overwrite what matters.

Public / private boundary

The framework is public-ready; private vault data is never published. The boundary is explicit, not implied.

Write policy

Skills may write only to allowed framework paths. Protected private paths are off-limits to agent writes.

Deny-publication patterns

Credentials, customer records, board materials, and personal contact details are blocked from crossing the boundary.

Human in the loop

Writes are dry-run by default; a human reviews and approves. The agent is a force multiplier, not an unsupervised actor.

Required checks

Validation runs before any push, so malformed or unciteable artifacts don't enter the record.

Provenance by default

Important claims carry source paths, dates, and confidence — the agent must cite.

Verifiable with GARL

MemexLab is GARL-ready: agent-authored changes can carry signed GARL Protocol receipts, so the provenance of an agent's commits is cryptographically checkable. (Receipt signing activates once the project is registered on GARL.)

Honest scope

What the agent layer is not

Not an autonomous researcher — a human approves writes.
Not a cloud memory service — the vault is local files you own.
Not a vector-database demo — retrieval is deterministic and citable.
Not production-stable — this is an early harness preview.

View on GitHub See it in the Showcase