Context Topics

Project continuity for one OpenClaw agent.

Context Topics lets one agent move between different bodies of work without mixing them together. Put on a product-launch hat, and the agent remembers the launch plan, source-of-truth files, decisions already made, and what should not be re-litigated. Switch to a client engagement, research thread, open-source project, or half-started idea, and that work gets its own room too.

It is meant to feel like project memory, not a prompt trick. You can start a topic intentionally, capture a conversation after you realize it matters, close the topic with durable notes, and come back later with the context still waiting.

Context Topics stores everything as plain files under ~/openclaw-soul/topics/ and uses OpenClaw's public plugin hooks. No gateway fork. No Control UI patch. No drifting away from upstream updates.

Why This Exists

One good agent can help with many things, but conversations blur together. Project A has different rules than Project B. Decisions made last week get forgotten. A useful chat becomes hard to resume because it was never promoted into a durable workspace.

Context Topics gives each project a small, file-backed memory room:

• a working pin for what the agent should know immediately
• topic-local memory
• decisions
• artifact index
• notes
• validation through /topic doctor

The goal is simple: one agent, many hats, each with its own continuity.

Quick Start

Install from ClawHub:

openclaw plugins install clawhub:openclaw-context-topics
openclaw gateway restart

Then open a chat and type:

/topic

Create a new topic:

/topic new product-launch

Load it later:

/topic product-launch

Close it when the work is done:

/topic close

The One-Minute Demo

You are chatting normally and realize the conversation should become a project:

/topic capture whatsapp-research

Context Topics creates:

~/openclaw-soul/topics/whatsapp-research/
  topic.md
  memory.md
  decisions.md
  artifacts/
    README.md
    index.md
  notes/

On the next agent turn, the agent is asked to fill the starter pin, memory, decisions, and artifact index from the current conversation.

Later:

/topic whatsapp-research

The agent gets the topic room back as working context.

Commands

Command	What it does
/topic	Show command help.
/topic list	Show available topic rooms.
/topic new <name>	Create a folder-backed topic room and activate it.
/topic capture <name>	Turn the current no-hat conversation into a new topic room.
/topic <name>	Load an existing topic into the current session.
/topic status	Show the active topic.
/topic panel [name]	Render a chat-stream topic panel.
/topic close [reason]	Run one final cleanup turn, update topic files, then clear the hat.
/topic refresh [name]	Ask the agent to refresh topic.md from topic memory and decisions.
/topic doctor [name]	Validate structure, pin quality, file references, probes, and review date.
/topic clear	Take the current hat off without cleanup.

Topic Room Shape

New topics use this folder shape:

topics/<name>/
  topic.md
  memory.md
  decisions.md
  artifacts/
    README.md
    index.md
  notes/

topic.md contains YAML code blocks for the always-loaded pin, file references, recent-memory rules, live probes, and metadata.

Topic-local files are intentionally plain Markdown so they can be read, edited, backed up, diffed, and committed like any other project artifact.

Example Pin

pin:
  title: "Product Launch"
  summary: "Launch planning for the next release."
  current_state:
    - "Launch date is not final."
    - "Pricing copy is still under review."
  operating_rules:
    - "Keep recommendations practical and release-oriented."
    - "Do not reopen settled positioning unless new evidence appears."
  settled_decisions:
    - "Use the existing landing page instead of creating a new microsite."
  open_work:
    - "Draft launch checklist."
    - "Confirm owner for customer announcement."
  avoid:
    - "Do not suggest paid ads until budget is approved."

Privacy And Safety

Context Topics is local-first.

• Topic rooms live under ~/openclaw-soul/topics/.
• Plugin session state lives under ~/openclaw-soul/state/.
• The plugin does not execute shell commands.
• Live probes are listed for the agent to run manually when needed.
• Sensitive-looking files are not inlined.
• Sensitive-looking absolute paths are redacted from prompt bundles.
• Session identifiers are hashed before persistence or logging.

You still control what goes into each topic. Do not put secrets, tokens, private keys, or credential material in topic memory.

Upstream Safety

This plugin uses OpenClaw's public plugin SDK:

• registerCommand for /topic
• before_prompt_build for per-turn context injection
• registerSessionExtension as an optional session metadata probe

It does not patch the gateway or Control UI bundle.

Feedback Wanted

This plugin is young. The most useful feedback is specific:

• Did /topic capture feel natural?
• Did /topic close write the right amount of memory?
• Did the topic pin help on the next session?
• Did anything feel too heavy, too magical, or too manual?
• What would make this safer for a team or shared machine?

Please open an issue with the command you ran, what you expected, what happened, and whether the topic files ended up useful.

Roadmap

Ideas being considered:

• richer file tiers and source-of-truth references
• better artifact indexing
• optional team/shared-topic conventions
• import/export of topic rooms
• stronger schema validation for topic manifests
• more polished chat-stream topic status

Links

• ClawHub: https://clawhub.ai/plugins/openclaw-context-topics
• Source: https://github.com/hussein1362/openclaw-context-topics

License

MIT-0. Free to use, modify, and redistribute.