Skip to content

V073-PM22B-LOCAL-AI-PROPOSALS.md

Latest commit

 

History

History
35 lines (24 loc) · 5.47 KB

V073-PM22B-LOCAL-AI-PROPOSALS.md

File metadata and controls

35 lines (24 loc) · 5.47 KB
Raw

PM2.2b — on-demand local AI memory proposals

Release: v0.7.3-pm22b-local-ai1, from operator-verified PM2.2a v0.7.3-pm22-scout-review1.

Scope and limitations

A new authenticated dashboard page Bob's Mind → AI ideas at /mind/memory/ai shows verified Scout sources, local generation status, untrusted AI proposals, and owner approval/rejection. It generates new concise memory ideas (observations, questions and tentative opinions) from a real selected, saved Moltbook Scout post. This is not generic reformatting of an old draft.

  • Generation is manual, never scheduled. After owner POST+CSRF, the server stores one queued job and starts one daemon thread after the database transaction commits, avoiding a blocked browser request.
  • The worker reads only the chosen post from the saved, checksum-verified Scout snapshot via the existing load_record verifier. It takes no Moltbook API credentials, and does not fetch any external content.
  • One bounded model call through Python stdlib HTTP client to 127.0.0.1:11434/api/generate (local Ollama), with optional BOB_MEMORY_OLLAMA_MODEL (default qwen3.5:9b). The model call has a 90-second network timeout, is non-streaming, bounded to 65,536 response bytes, and is asked for strict JSON and 0–3 proposals. It has no OpenClaw agents, tools, browser, or publisher capabilities.
  • Returned candidates must contain a source post ID exactly matching the verified post and a small schema of title, note, category, low/medium confidence and rationale. Overlong, malformed, contradictory source IDs, unsupported confidence/status, or some obvious instructions/prompt tokens fail the entire batch. This is not a guarantee of factual truth or a full defense against prompt injection: all proposed text remains untrusted and must be reviewed.
  • One active job at a time (database unique partial index), three generation attempts per Scout run, max 3 proposals/job, 30 pending / 120 total proposals. Exact deduplication of normalized same-run proposal content, persistent state queued/running/completed/failed, and friendly status text. Interrupted stale jobs are marked failed, not silently retried.
  • Approval is always manual. Owner can edit the candidate title/text, category and confidence and must supply a reason. Only then does the reviewed proposal become a new private PM2.1 memory. Source attribution is server-owned (Scout run and post ID) and reverified before acceptance. Rejection creates no memory. Review decisions and source fingerprints stay local.
  • Failed generations store only fixed error codes. Raw model output, host exception tracebacks, private directories and credentials never appear in dashboard error pages.
  • Data are in additive private SQLite tables bob_memory_ai_jobs and bob_memory_ai_proposals; no change to existing memory records/PM2.2a tables or publication state. No auto-recall, SOUL.md/IDENTITY.md editing, scheduled reflection, publishing or outbound Moltbook activity.

Runtime/network prerequisite

Ollama must be reachable from OpenClawLab WSL2 on its loopback interface 127.0.0.1:11434, with model qwen3.5:9b available (or a model set in the Activity Manager's process environment via BOB_MEMORY_OLLAMA_MODEL). This is not the same test as Ollama being available from Windows. If WSL networking does not forward loopback, jobs will display Local Ollama could not complete the request. Do not expose Ollama publicly to make this work. Any explicit trusted-LAN host configuration belongs in a separate hardened connectivity change.

The Activity Manager itself currently uses HTTP Basic on a trusted LAN and is not HTTPS. Memories may be personal: keep it private behind VPN or a TLS reverse proxy, never public Internet.

Acceptance

  1. From WSL OpenClawLab run bobctl update. Confirm the exact path v0.7.3-pm22-scout-review1 to v0.7.3-pm22b-local-ai1 (three installed files) and approve with UPDATE. Activity Manager restart only.
  2. Visit Bob's Mind → AI ideas and choose one previously verified Scout post. Click Ask local Ollama for memory ideas. The request should promptly redirect to the same page with a queued/running status. Refresh later to see completed or failed status; there is no fake percentage.
  3. If completed, inspect 0–3 proposals. Model-generated text is marked unverified, with source run ID/post ID, rationale and confidence. If Ollama is not accessible, the page reports an accurate fixed failure status; it must not hang indefinitely.
  4. Edit a proposal for accuracy, enter a reason, and approve. It appears as a new private memory with source attribution. Reject a second proposal and verify it is not added.
  5. Attempt to resubmit a stale review form: the server must refuse a second memory. Check Scout & review, Bob's Mind → Suggestions, Personality and existing queued posts are unchanged. No new Moltbook post is authorized or generated.
  6. Verify the new job can recover from process restarts without repeating API requests; stale jobs become failed, and no auto-retry occurs.

Follow-ons

PM2.2b is operator-triggered. For truly autonomous but safely permissioned suggestion scheduling, build a separate PM2.2c job worker with model/publisher privilege separation, source privacy retention/export, additional evidence validation and explicit opt-in. PM2.3 remains explicitly opt-in bounded read-only recall of owner-approved active memories, never control prompts or tool instructions. FM1 Friends Manager comes after PM2 acceptance.