OpenClaw

Install PowerContext in OpenClaw’s memory slot, then verify its five tools, private-session gate, Scope, and Server boundary.

OpenClaw support is Version-gated at >=2026.8.1-beta.2. Node 24.15+ is recommended for that supported host release. The plugin package declares a broader Node >=20 engine, but that is not the same as the recommended host environment.

OpenClaw gets a deliberately narrow integration: automatic Memory recall, bounded Source capture, and five explicit Memory tools. It does not expose Handoff, Experience, Candidate Review, or Skill operations.

Install and activate

Install PowerContext and the OpenClaw plugin:

uv tool install --force "powercontext[cli,server] @ git+https://github.com/oceanbase/powercontext.git@master"

powercontext setup openclaw \
  --source oceanbase/powercontext \
  --ref master

powercontext server run

openclaw

Setup requires openclaw and pnpm. It checks the OpenClaw version, installs dependencies with the frozen lockfile, builds dist/index.js, links the plugin, enables it, selects the memory slot, appends the five tools to tools.alsoAllow, enables recall/capture, sets gateway.mode=local when absent, and restarts the Gateway.

It does not start PowerContext Server. The command sequence shows that boundary plainly: setup first, Server separately, then the host.

To use project mode or a different endpoint:

powercontext setup openclaw \
  --server-url http://127.0.0.1:9000 \
  --scope-mode project

Runtime defaults

Run once with defaults before changing them. This gives you a known local baseline for separating host configuration from Server or authentication failures.

Key	Default	Meaning
endpoint	written by setup	PowerContext Server URL
tokenEnv	POWERCONTEXT_CLIENT_API_TOKEN	Name of the Gateway environment variable containing the bare token
timeoutMs	2500	Request timeout, allowed range 250–15000
prepareMaxBytes	8000	PreparedContext budget, allowed range 512–32768
autoRecall	true	Prepare before prompt construction
autoCapture	true	Capture after a successful agent end
captureMaxChars	4000	Characters in the formatted User: ... Source payload, including its prefix; allowed range 128–20000
scopeMode	agent	agent or project

CLI setup rejects endpoint credentials, query strings, and fragments. Manual plugin configuration is less strict, so use the setup command or apply the same operator checks yourself.

Exactly five tools

OpenClaw exposes this exact Memory subset. If the task needs Handoff, Review, Experience, or Skill, choose another integration instead of trying to infer hidden tools.

Tool	Operation
powercontext_memory_search	Search durable Memory; session transcripts and wiki are not searched
powercontext_memory_get	Read one exact encoded PowerContext citation
powercontext_memory_store	Store a Memory entry
powercontext_memory_revise	Revise one exact current citation
powercontext_memory_retire	Retire one exact current citation

Search accepts memory, wiki, all, or sessions as a host corpus hint, but wiki is unsupported and sessions does not search transcripts. all still means durable Memory here.

Get defaults to 120 lines. It truncates the Memory body at 12,000 characters, then may append a continuation notice, so final tool text can be slightly longer. The caller may request more than 120 lines, so 120 is not a hard maximum. Store and revise accept at most 8192 UTF-8 bytes at runtime. A stale revise returns a conflict with the action: search again and retry using the current exact citation.

These five tools are the whole PowerContext surface in OpenClaw. There is no Handoff, Review, Experience, external Skill, managed Skill, or export tool in this plugin.

Minimal verification

Run Server and host diagnostics separately:

powercontext doctor
powercontext ready
powercontext capabilities
powercontext doctor openclaw

Then use a direct/private OpenClaw session to:

1. Store a harmless test preference with powercontext_memory_store.
2. Search for it with powercontext_memory_search.
3. Read the returned citation with powercontext_memory_get.
4. Revise it, then retry the old citation and confirm the conflict.
5. Retire the current citation and confirm it no longer appears as active Memory.

Also run the learning-path question used by the other hosts. First store it with powercontext_memory_store:

Use PowerContext to remember this decision:
after changing the OpenAPI contract, regenerate the client and inspect the diff.
Return the Memory Citation. Do not store secrets.

Then ask:

What should happen after the OpenAPI contract changes?

doctor openclaw only checks that the CLI exists and that memory-powercontext reports enabled=true, status="loaded", and memorySlotSelected=true. It does not check OpenClaw version, endpoint, Gateway token, Server reachability, the five allowlist entries, or an actual recall/capture round trip. Single-host setup does not run this final doctor for you.

Private-session gate

Recall, capture, and tools run only for eligible private sessions:

• Explicit chatType=group and chatType=channel are rejected.
• If chatType is present, only direct is accepted.
• Incognito session keys are rejected.
• Session keys parsed as group or channel are rejected.
• If metadata lookup throws, eligibility is rejected. If lookup only returns no entry, the plugin falls back to session-key and chat-type heuristics; an opaque non-group, non-channel, non-incognito key may still pass.

A blank session key can still be accepted. “Private” is an eligibility classification, not secret sanitization. The plugin does not redact secrets from an otherwise eligible direct prompt.

Capture and recall are bounded

Before prompt construction, the plugin sends only the latest user text as the prepare query, capped at 8192 UTF-8 bytes. Returned content must match the PreparedContext schema. Ready content is wrapped as untrusted history, and recalled boundary tags are escaped before injection.

After a successful agent_end, capture stores only the latest eligible user message:

User: <text>

It does not capture the assistant answer or synchronize the full transcript. Metadata records the raw agent ID, an optional hashed session identifier, an optional channel/provider, and privacy_class: private.

Before compaction and at session end, the plugin checks capabilities and Flushes observed scopes only when memory_extraction=true. If extraction is disabled, captured Sources remain pending.

Scope is still per agent

Agent mode resolves to:

openclaw:agent:<URL-encoded-agent-id>

Project mode resolves to:

openclaw:agent:<URL-encoded-agent-id>:project:<first-32-hex-of-sha256-project-key>

Project mode therefore narrows Memory by both agent and project. It does not create a shared cross-agent project namespace. With zero or multiple trusted active project keys, the plugin falls back to agent scope.

Gateway token and remote HTTP

When Server auth is enabled, put the same bare token in the Gateway process environment:

export POWERCONTEXT_SERVER_AUTH_ENABLED=true
export POWERCONTEXT_SERVER_AUTH_TOKEN="$POWERCONTEXT_LOCAL_TOKEN"

powercontext server run

Gateway environment:

printf '%s\n' 'POWERCONTEXT_CLIENT_API_TOKEN=<same token value>' >> ~/.openclaw/.env
chmod 600 ~/.openclaw/.env
openclaw gateway restart

The plugin reads the variable named by tokenEnv and adds the Bearer header. Exporting the token only in the shell that started PowerContext Server does not put it in the Gateway environment.

For a non-loopback endpoint, HTTPS is an operator requirement. Current setup and plugin code do not enforce it. A token does not make plaintext HTTP safe.

Execution flow

Lifecycle recall, capture, and Flush fail open. Explicit tool failures return structured unavailable, rejected, or conflict results. Unlike Hermes, the OpenClaw HTTP client uses standard fetch() behavior; the documented behavior does not claim that redirects are refused.

Negative paths to test

• Stop PowerContext Server. The ordinary OpenClaw response should continue without recall; an explicit store must report unavailable.
• Open a group, channel, or incognito session. Recall, capture, and all five tools must be unavailable there.
• Put the same project key under two different agent IDs. Their project-mode Scopes must remain different.
• Remove the token from the Gateway environment while Server auth remains enabled. Doctor may still pass, but live operations must fail visibly.
• Reuse an old citation after revise. The operation must return conflict rather than overwrite the current Revision.
• Capture a user turn with an assistant response. Only the latest user prompt should be sent as Source content.
• Configure a remote http:// endpoint. Treat this as an unsafe operator configuration; the plugin will not block it for you.

Current limits

• Only the Memory plane is supported. Handoff, Review, Experience, Skill, external Skill, and managed export are Unsupported here.
• Project Scope still includes agent ID; cross-agent sharing is not implemented.
• Capture does not include a full transcript or assistant output, and it does not perform secret redaction.
• Setup is not transactional; a late failure may leave partial host changes.
• Reusing a cached checkout for a mutable ref may not fetch its latest commit.
• Doctor is an installation check, not a Server or runtime E2E test.
• A real OpenClaw host-process E2E is Not validated here.

Finish the shared check on this host

What should happen after the OpenAPI contract changes?

1. Store the answer in a stable Scope with powercontext_memory_store. Of the five Memory tools, only store performs this write.
2. Open a new private / direct session and ask the same question.
3. The recall must include a Citation and enter the model only as untrusted history from this latest-user prepare.
4. Stop the Server and ask again. Automatic recall may fail. The session should continue.
5. Call powercontext_memory_store again. The failure must stay visible. Do not pretend the write succeeded.

Completion checklist

• OpenClaw is >=2026.8.1-beta.2; the runtime uses the recommended Node 24.15+ environment.
• pnpm is installed, setup completed, and the Gateway restarted.
• PowerContext Server was started separately.
• doctor openclaw passes, with its boundary understood.
• All five exact tool names are allowlisted and usable in a private/direct session.
• Group, channel, and incognito negative tests pass.
• Capture contains only the latest user prompt, not the transcript.
• Agent and project Scope formulas match the intended isolation; no cross-agent sharing is assumed.
• The token exists in the Gateway environment.
• Every remote endpoint uses HTTPS by operator policy.

For a broader tool surface, use Hermes. For governed Skill export, continue with Let the agent extend itself.