Continuity Memory Contract Plus (CMCP) for OpenClaw

Looking for the current project? Start with CMCP-TIME. This repository preserves the earlier OpenClaw-oriented policy contract and review artifact. Current development continues separately in CMCP-TIME, with source-preserving history, original timestamps, bounded reading and user-controlled follow-ups. The documentation below describes this historical bundle. Its OpenClaw assumptions, verification claims and submission guidance do not define CMCP-TIME. This notice does not change existing code, releases or licenses, and is not an automatic migration guide.

目前主線請見 CMCP-TIME。 此倉庫保留早期以 OpenClaw 為對象的政策契約與審閱成果；下方是該歷史版本的說明，不代表新版的功能、授權或相容性。三個倉庫分開保留，不會自動搬移使用者資料。

CMCP is a policy-first continuity memory bundle for OpenClaw.

It is designed for builders who want long-lived agent behavior without leaving memory semantics to prompt drift, UI copy, storage heuristics, or host-local guesswork.

Concrete implementation reference: OpenClaw Continuity is the usable OpenClaw skill package that demonstrates the product path behind this policy contract.

The bundle includes:

• CMCP Core: the canonical continuity memory contract
• CMCP Guard: runtime enforcement for write decisions, correction flows, and /new
• CMCP Setup: the disclosed capture model for onboarding/setup
• CMCP Policy: the inspection and manual-control surface model

What Problem CMCP Solves

CMCP makes one shared policy explicit across memory layers and product surfaces.

It answers:

• what stays session-only
• what becomes staged continuity
• what is promoted to tracked follow-up
• what belongs in daily writeback only
• what may become durable personalization
• what must never persist
• how /new should carry continuity forward
• how onboarding may create initial durable state
• how user correction overrides prior memory

The core posture is strict:

• ordinary chat is read-only by default
• persistent writes require explicit authorized invocation areas
• restricted personalization requires disclosed capture and explicit user ownership
• forbidden content is rejected rather than silently persisted
• host-provided storage and host-provided /new inputs are treated as untrusted

What Is In This Repository

.codex-plugin/   Plugin metadata
docs/            Audit, positioning, proposal, and publish docs
hooks/           CMCP Guard hook entry
scripts/         Acceptance, adversarial, end-to-end, and integration verification
skills/          CMCP Core contract and schemas
src/             Runtime policy, storage, correction, and safety logic

Key entry points:

• Contract: skills/cmcp-core/contract/cmcp-contract-v1.yaml
• Runtime policy: skills/cmcp-core/contract/cmcp-runtime-policy.json
• Host integration boundary: skills/cmcp-core/contract/cmcp-host-integration-v1.yaml
• Guard hook: hooks/cmcp-guard/HOOK.md
• Audit summary: docs/CMCP_AUDIT_SUMMARY.md
• Review packet: docs/REVIEW_PACKET.md

Relationship To OpenClaw Continuity

CMCP is the policy contract and review artifact.

OpenClaw Continuity is the separate product skill package that demonstrates the same continuity problem in a usable OpenClaw implementation: /new carryover, staged memory, tracked follow-up, time-aware care, natural-language settings, closure, cooldown, and structured writeback.

For OpenAI or Anthropic review, treat CMCP as the primary artifact and OpenClaw Continuity as implementation evidence. The two should not be merged into one review package.

Current Product State

CMCP is currently strongest as:

• a technical review artifact
• an integration baseline for OpenClaw hosts
• a policy contract with executable enforcement
• a failure-path-verifiable memory layer

CMCP is not currently positioned as:

• a polished end-user UI product
• a turnkey installer UX
• a semantic classifier for all risky inferences
• a distributed storage system

Contact and collaboration / 聯絡與合作

Interested in this project? I'd love to hear your ideas, feedback, and collaboration proposals, including commercial partnerships. Feel free to reach me at either address:

• adarobot666@gmail.com
• redwakame616@gmail.com

If you find this project useful, a GitHub star is a welcome show of support. I'll continue improving and updating my open-source projects, and your feedback helps shape what comes next.

如果你對本專案有興趣，歡迎透過上述任一信箱交流想法、分享使用回饋，或洽談合作與商業應用。

如果這個專案對你有幫助，也歡迎在 GitHub 給一顆 Star，作為支持與鼓勵。我會持續優化與更新開源專案，並將大家的回饋作為後續改進的參考。

Verification

CMCP includes executable verification, not only prose.

Available scripts:

npm run check
npm run acceptance
npm run adversarial
npm run e2e
npm run pipeline
npm run host
npm run schema
npm run demo:host
npm run matrix

The current verification layers cover:

• intended product behavior
• adversarial failure paths
• end-to-end integration flows
• host integration behavior
• runtime shape validation
• multi-host and multi-modal wiring checks

Install / Run Locally

Requirements:

• Node.js 18 or newer

Run:

npm install
npm run check
npm run adversarial

If you want a concrete runnable example, use:

npm run demo:host

Generated demo artifacts appear under:

• runtime-data/mock-host-demo/

Recommended Reading Order

For a reviewer who wants the shortest correct path:

1. docs/REVIEW_PACKET.md
2. docs/CMCP_AUDIT_SUMMARY.md
3. skills/cmcp-core/contract/cmcp-contract-v1.yaml
4. hooks/cmcp-guard/HOOK.md
5. docs/openclaw-cmcp-proposal-en.md or another target-specific proposal

For a host integrator:

1. skills/cmcp-core/contract/cmcp-host-integration-v1.yaml
2. src/cmcp-guard/cmcp-storage-adapter.js
3. scripts/run-cmcp-host-integration-smoke.mjs
4. scripts/run-cmcp-integration-matrix.mjs

Docs

• Claude review handoff: docs/CLAUDE_REVIEW_HANDOFF.md
• Review packet: docs/REVIEW_PACKET.md
• Audit summary: docs/CMCP_AUDIT_SUMMARY.md
• Requirements coverage: docs/REQUIREMENTS_COVERAGE.md
• OpenAI / Anthropic send list: docs/OPENAI_ANTHROPIC_SEND_LIST.md
• GitHub prep checklist: docs/GITHUB_PREP_CHECKLIST.md
• Bug disposition: docs/BUG_DISPOSITION.md
• Outreach email templates: docs/OUTREACH_EMAIL_TEMPLATES.md
• Official submission messages: docs/OFFICIAL_SUBMISSION_MESSAGES.md
• Product positioning: docs/PRODUCT_POSITIONING.md
• Publish plan: docs/PUBLISH_PLAN.md
• OpenAI proposal: docs/openai-cmcp-proposal-en.md
• Anthropic proposal: docs/anthropic-cmcp-proposal-en.md
• OpenClaw proposal: docs/openclaw-cmcp-proposal-en.md

Safe Claims

These claims are supportable by the current artifact:

• CMCP is policy-first continuity memory for OpenClaw
• the bundle includes executable adversarial and end-to-end verification
• persistent memory is not ambient by default
• forbidden content is rejected rather than silently persisted
• correction flows are fail-closed
• /new uses structured continuity rather than transcript replay

Deliberately Deferred Boundaries

CMCP does not currently claim:

• automatic semantic detection of all inference-forbidden categories
• distributed storage guarantees
• full compliance certification
• elimination of every possible leak in arbitrary hosts

Those boundaries are intentionally documented rather than hand-waved away.