openclaw-context-pack 1.9.20

Local-first AI agent memory governance: Store / Pack / Observe with citations, receipts, and rollback

openclaw-mem

Local-first context supply chain for AI agents — Store / Pack / Observe with citations, trust-policy receipts, and rollback.

openclaw-mem captures agent activity as durable local records, then assembles bounded ContextPack bundles with inspectable include/exclude reasons and citation coverage. Use it first as a plain SQLite sidecar; promote to deeper OpenClaw integration only after the local proof earns the extra surface.

Start here

1. Run the synthetic proof: Trust-policy synthetic proof
2. Pick an evaluation path: 5 minutes / 30 minutes / one afternoon
3. Check Core vs Advanced Labs: Core vs Advanced Labs
4. Choose sidecar vs engine: Install modes
5. Check shipped vs partial status: Reality check & status
6. Read in Traditional Chinese: Traditional Chinese edition

What is automatic today?

Surface	Status	Meaning
Sidecar observation capture	Automatic when the plugin is enabled	Captures JSONL observations with backend/action annotations.
Harvest, triage, and graph capture	Scheduled on configured hosts	Converts captured records into searchable stores and receipts.
pack	CLI core	Produces bounded ContextPack output with citations and trace receipts.
Harness-persistent memory	Opt-in via Gateway + CLI install card	Installs a persistent Codex / Claude / Gemini / generic agent memory posture backed by capability-scoped gateway tokens and parity-aware read diagnostics.
Graph routing, optimize assist, continuity, GBrain	Advanced Labs / opt-in lanes	Available for mature operators, but not part of the first evaluation path.
Mem-engine Proactive Pack	Optional promotion	Bounded pre-reply recall orchestration after explicit engine adoption.

Why it is different

• Governance before recall theater — trust policy, citations, trace receipts, and rollback posture make memory inclusion inspectable instead of opaque.
• Local-first by default — JSONL + SQLite, no external database required.
• Cheap recall loop — search → timeline → get keeps routine lookups fast and inspectable.
• Bounded packing — pack emits a stable ContextPack contract for injection, citations, trust-policy receipts, and trace-backed debugging.
• Symbolic canvas helper — symbolic-canvas can turn task traces into compact Mermaid + node_id drill-down artifacts without live capture, Gateway patching, or memory mutation.
• Fits real OpenClaw ops — capture tool outcomes, retain receipts, sanitize runtime artifacts, and keep rollback simple.
• Upgradeable path — sidecar first, engine later; no forced migration on day one.
• Harness-friendly — external coding agents can get a persistent memory posture through the Memory Gateway without receiving raw SQLite or workspace-file authority. Parity-aware gateway reads can search/pack configured workspace Markdown memory and report when a no-result answer is only partial.
• Advanced labs are opt-in — graph routing, GBrain, continuity, Dream Lite, Self Curator review packets, and deeper optimization lanes stay out of the first evaluation path.

Why this exists

Long-running agents do not just forget. They also accumulate memory that quietly degrades:

• old notes still match the query even when they are no longer useful
• untrusted or hostile content can retrieve well and slip into context
• prompts bloat into giant memory dumps instead of a small, inspectable bundle
• when something goes wrong, it is hard to explain why a memory was included

openclaw-mem tackles that by building compact memory packs with citations, trace receipts, and trust-policy controls.

Try it in 5 minutes

You can prove the core behavior locally without touching OpenClaw config.

For the packaged CLI:

python -m venv .venv
. .venv/bin/activate
pip install openclaw-context-pack
openclaw-mem --db /tmp/openclaw-mem-demo.sqlite status --json

For the repository proof fixture:

git clone https://github.com/phenomenoner/openclaw-mem.git
cd openclaw-mem
uv sync --locked

uv run --python 3.13 --frozen -- \
  python benchmarks/trust_policy_synthetic_proof.py --json

What this proof shows

• vanilla packing selects a quarantined row from synthetic memory
• trust-aware packing excludes that row with an explicit reason
• selected rows keep citation coverage and traceable receipts

Full proof path:

• Evaluator path
• Trust-policy synthetic proof
• Trust-aware pack proof
• Command-aware compaction proof
• Metrics JSON
• Synthetic fixture + receipts
• Inside-out demo

Store + Pack + Observe

The product loop is simple and stable:

1. Store: capture, ingest, and query observations with store/ingest/search.
2. Pack: run pack to get a bounded bundle_text and context_pack (schema: openclaw-mem.context-pack.v1), with citations, trust policy, and trace receipts.
3. Observe: use timeline, get, and artifact outputs for explainability and rollback.

When mem-engine is active, Proactive Pack extends the same Pack contract into live turns as a small, receipt-backed pre-reply bundle.

For external AI harnesses, openclaw-mem harness install writes a managed persistent instruction card for Codex, Claude, Gemini, or a generic agent surface. The card points the harness at OPENCLAW_MEM_GATEWAY_URL / OPENCLAW_MEM_GATEWAY_TOKEN; raw tokens are never written into prompt files. Capability-scoped tokens let operators choose read-only, proposal/append write, admin, or owner-equivalent direct-store authority. For Codex specifically, openclaw-mem codex install adds a stronger Superpowers-style surface with generated CLI shims, env/gateway checks, and codex doctor verification.

Read more: Harness-persistent memory install and the release notes for version-specific gateway behavior.

Advanced labs

The first-time evaluator path is Store / Pack / Observe. Everything below is opt-in after the core proof is clear.

Advanced lanes currently include:

• Graph routing for topology-aware recall experiments.
• GBrain sidecar for bounded read-only lookup and restricted helper-job experiments.
• Governed continuity side-car for derived continuity inspection and public-safe summaries.
• Dream Lite / deeper optimize loops for research-grade memory maintenance workflows.
• Self Curator engine for checkpointed skill lifecycle review/apply loops. The current scheduled lane can mutate SKILL.md body sections with rollback receipts; memory/dream/authority expansion remains gated behind explicit review.

These lanes are not required for the 5-minute proof, the sidecar install path, or the basic ContextPack contract. Treat them as labs until your use case needs them.

Read more:

• Product positioning
• Compounding Context OS roadmap
• Core vs Advanced Labs
• Evaluator path
• Architecture
• Context pack
• Symbolic Canvas
• Symbolic canvas ops skill card
• Experimental GBrain sidecar
• Optional Mem Engine

OpenClaw 2026.4.15 and openclaw-mem

By OpenClaw 2026.4.15, the native memory and prompt-time integration experience had become noticeably stronger. We are genuinely happy to see that direction mature.

That is good for the ecosystem, good for operators, and good for openclaw-mem too. A stronger foundation makes it easier to keep our own work focused on what matters most: better packs, clearer evidence, and safer memory maintenance.

Our direction is not to shrink back into native features. It is to build a clearer, more opinionated product layer on top of a stronger foundation.

Read more:

• Why openclaw-mem still exists in a stronger OpenClaw world
• openclaw-mem and OpenClaw 2026.4.15

Deeper operations live below the fold

openclaw-mem also has governed memory-hygiene and artifact-observation tools for mature operator stacks. They are useful after the core product is proven, but they are not required for the first evaluation path.

Start with:

• Core vs Advanced Labs
• Evaluator path
• Governed optimize assist
• Hermes Curator adoption review
• Self Curator v0 review-only sidecar contract
• Command-aware compaction proof

Manual Self Curator scout smoke:

openclaw-mem self-curator skill-review \
  --skill-root ~/.openclaw/workspace/skills \
  --out-root .state/self-curator/runs \
  --json

Checkpointed Self Curator apply flow:

openclaw-mem self-curator plan --mutations-file mutations.json --out plan.json --workspace-root . --json
openclaw-mem self-curator apply --plan plan.json --workspace-root . --checkpoint-root .state/self-curator/checkpoints --receipt-root .state/self-curator/apply-runs --json
openclaw-mem self-curator verify --receipt .state/self-curator/apply-runs/<run>/apply-receipt.json --json
openclaw-mem self-curator rollback --receipt .state/self-curator/apply-runs/<run>/apply-receipt.json --json

The scout emits review-only lifecycle artifacts. The apply flow may mutate whitelisted relative workspace files through explicit plans, checkpoints, diffs, receipts, verification, and rollback.

Autonomy controller / cron posture:

python3 /root/.openclaw/workspace/openclaw-mem/tools/self_curator_controller.py \
  --repo /root/.openclaw/workspace/openclaw-mem \
  --workspace-root /root/.openclaw/workspace \
  --skill-root /root/.openclaw/workspace/skills \
  --out-root /root/.openclaw/workspace/.state/self-curator/controller-runs \
  --mode unattended_apply \
  --max-mutations 5 \
  --cron-output

Scheduled operation should use a separate cron-runner job, not heartbeat. Unattended apply is 事後報備制: the controller applies whitelisted low-risk changes first, emits NEEDS_CK with report/rollback paths when changes happen, and rollback is available from the apply receipt. The scheduled policy mutates SKILL.md directly by appending a bounded ## Curator lifecycle section rather than writing detached metadata only.

More links

Core and adoption

• Why openclaw-mem still exists: docs/why-openclaw-mem-still-exists.md
• OpenClaw 2026.4.15 comparison: docs/openclaw-2026-4-15-comparison.md
• About the product: docs/about.md
• Proactive Pack: docs/proactive-pack.md
• Choose an install path: docs/install-modes.md
• Detailed quickstart: QUICKSTART.md
• Docs site: https://phenomenoner.github.io/openclaw-mem/
• Traditional Chinese edition: docs/zh/index.md
• Reality check / status: docs/reality-check.md
• Deployment patterns: docs/deployment.md
• Auto-capture plugin: docs/auto-capture.md
• Agent memory skill (SOP): docs/agent-memory-skill.md
• Optional Mem Engine: docs/mem-engine.md
• Release notes: https://github.com/phenomenoner/openclaw-mem/releases

License

Dual-licensed: MIT OR Apache-2.0. See LICENSE, LICENSE-MIT, and LICENSE-APACHE.