docs(plans): mem0-v3-locomo activation plan + NEEDS_APPROVAL on license

[OpenCircuitDev]

Summary

Track 2, item 1 from docs/AGENT_OPERATIONS.md — plan-only deliverable for activating the INACTIVE mem0-v3-locomo sandbox. No code. The plan is at docs/superpowers/plans/2026-06-11-mem0-v3-locomo-activation-plan.md.

Plan scope (5 sections per the Task 3 directive)

1. LoCoMo dataset acquisition — repo, file, size, license, access requirements.
2. Harness design — files + flow, reusing the patterns from mem0-library-retrieval-recall (feat(bench): flip mem0-library-retrieval-recall to ACTIVE + discovered upstream bug #56, just merged) and amnesia-ab.
3. Hardware/runtime on the operator's dev box (Ollama llama3 8B Q4).
4. Mapping onto the locked expected.json contract — ≥88 confirm, <80 refute.
5. Wall-time + blockers.

Plus 7 bite-sized tasks (Task 0 = verify upstream, Task 7 = run the verdict).

NEEDS_APPROVAL — operator decision required before Tasks 2+ can run

LoCoMo ships under Creative Commons Attribution-NonCommercial 4.0 International (verified by API fetch of LICENSE.txt on snap-research/locomo@main). The OCM repo is Apache 2.0. This intersection is the only spec-touch decision and is flagged as a 4-option multi-choice in §1 of the plan.

The plan-side default is (a) fetch-on-run, no bundle — keeps the repo's Apache 2.0 file surface clean and the dataset never ships inside OCM. SHA-pinned URL for reproducibility. ~3 MB per fresh-cache run.

Operator can pick (b) bundle with attribution, (c) synthetic proxy, or (d) defer. Switching costs are documented per-option.

Key facts the plan locked down

Aspect	Value	Notes
Dataset	data/locomo10.json on snap-research/locomo	10 conversations × ~300 turns × 9K tokens avg
Size	2.81 MB (2,805,274 bytes)	Single HTTPS GET; NOT >1GB
Auth	None	No account, no API key, no payment
License	CC BY-NC 4.0	The one operator decision
Embedder	sentence-transformers/all-MiniLM-L6-v2	Same as #56's hermetic config
Vector store	faiss-cpu	Same as #56
Mem0 mode	add(infer=False) + search(filters={"user_id": ...}, top_k=10)	Same as #56
LLM (--with-llm only)	Ollama llama3 8B Q4 via host.docker.internal:11434	Same as amnesia-ab
Verdict run wall-time	~10-15 min for the 3-repeat run	Under the 1800s timeout_seconds
--with-llm diagnostic	OFF by default; ~30-50 min when on	Operator-asked-for E2E path, kept out of the contract metric

Mapping onto the contract (no expected.json changes)

Field	Where the value comes from
primary_value (locomo_recall_score)	mean per-conversation recall × 100 (0-100 scale matching the published 91.6)
secondary_value (tokens_retrieved_p50)	True per-QA p50 across all questions (Task 5 fixes the naive per-conv approximation)
status flip INACTIVE → ACTIVE	Task 6, in the same commit that lands bench.py + docker-compose.yml
blocked_on: ["...MCP local mode not yet packaged..."]	Reframed in README — that's a daemon concern, not a verdict blocker; this sandbox tests the library-driven retrieval pattern (Mem0's Memory.from_config(), same as #56)

Test plan

This PR is documentation. No CI tests apply. The plan's own tasks (when run later) follow TDD as written.

• Plan covers all 5 sections of the directive (self-review checklist at end of doc)
• LoCoMo dataset facts verified by live API fetch against snap-research/locomo
• License determined by fetching + base64-decoding LICENSE.txt from the repo
• Patterns from feat(bench): flip mem0-library-retrieval-recall to ACTIVE + discovered upstream bug #56 + amnesia-ab cited with file paths
• expected.json contract field-by-field mapping
• Wall-time estimate broken down per phase
• NEEDS_APPROVAL formatted as multi-choice per directive
• Operator picks (a/b/c/d) on Q1 ← unblocks Tasks 2-7

What this PR does NOT do

• No code changes (consistent with "plan only, no execution" in the directive).
• No expected.json changes — the contract is already locked.
• No dataset download — that happens in Task 2 only after Q1 is resolved.

🤖 Generated with Claude Code