The cheapest discriminating test of the central loop (spec row 9, library-
driven retrieval) as a faithful miniature: mxbai-embed-large cosine top-5 ->
inject -> llama3 8B Q4 via Ollama, 62-memory corpus w/ cross-project
distractors, 20 tasks, objective key-fact scoring.

Measured (results/run-2026-06-11T20-32-21.json):
  memory_on_fact_recall_pct  94.2  (confirm >=70)
  retrieval_hit_rate_pct    100.0  (confirm >=80)
  memory_off_fact_recall_pct  2.5  (sanity <=25 — corpus not guessable)
  latency p50 on/off         19.5s / 12.3s

OFF-arm failure mode is confident fabrication, not ignorance — the memory
loop is the difference between correct specifics and plausible lies on an
8B model. Per the decision rule this justifies: the Ollama backend adapter,
activating mem0-v3-locomo, and cutting v0.1.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… Ollama users

Third InferenceBackend: bridges OCM to an existing Ollama daemon via its
native NDJSON /api/chat API (model tag required per-request; max_tokens
maps to options.num_predict; health via /api/tags). Selector untouched —
explicit construction for now; daemon settings wiring is the follow-up.

Parser test fixtures are VERBATIM captures from a live Ollama daemon
(llama3, 2026-06-11) — pinned to the real wire format. Motivated by the
amnesia-ab sandbox CONFIRMED verdict (94.2% fact recall on this exact
daemon + model class).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… blocker resolved

The 'five model SHA256 hashes' pre-release blocker was cleared when the
unhashed Qwen3 entries were dropped (#50); the shipping registry is 3
models, all hashed. README still claimed 5 GGUFs + an open blocker.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ose + bench.py

The dry-run validator requires docker-compose.yml + bench.py for ACTIVE
sandboxes (caught by Bench Framework CI on PR #67 — the framework doing
its job). bench.py delegates to run.mjs (ONE harness, the exact artifact
that produced the CONFIRMED result); compose runs it in node:22-slim
against the HOST Ollama daemon via host-gateway, same host-dependency
pattern as vllm-q4-llama8b. run.mjs now honors OLLAMA_URL.

Local validate_compose: PASS.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…row 9

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>