[yusiwen]

What changed

• docs/memory-bench.md: the measured table replaces the empty cell. Two models, 2026-10-26, local
  Ollama 0.40.11, each run twice and byte-identical both times; the raw MEMBENCH lines are quoted
  under the table so the numbers can be re-derived from the page itself.
• docs/roadmap.md: B3's Delivered note no longer says the measurement does not exist; it states the
  result (0/6 → 4/6 and 3/6 at hit@3 on paraphrases; 0/4 → 1/4 and 4/4 cross-lingual).
• agent/memory_bench_test.go: the gated run now logs the vector dimension, since a vector's shape is
  part of what was measured.

The numbers

Ranker	Model	Group	n	hit@1	hit@3	hit@5	MRR
lexical	—	paraphrase	6	0.00	0.00	0.00	0.00
semantic	nomic-embed-text (768)	paraphrase	6	0.33	0.67	0.83	0.49
semantic	bge-m3 (1024)	paraphrase	6	0.50	0.50	0.83	0.61
lexical	—	crosslingual	4	0.00	0.00	0.00	0.00
semantic	nomic-embed-text (768)	crosslingual	4	0.00	0.25	0.25	0.08
semantic	bge-m3 (1024)	crosslingual	4	0.75	1.00	1.00	0.83

Both models keep hit@1 = 1.00 on the word-sharing group, so the semantic path did not trade the easy
case away.

Why

This is the measurement the previous PR deliberately left open. It also separates two claims that were
being made together: paraphrase recall is a property of having embeddings at all (either model closes
it), while cross-lingual recall is a property of which model — the English-focused one finds one of
four, the multilingual one finds four of four.

Verification

• go test ./... -count=1 → 16 packages ok; gofmt and go vet clean on the changed file.
• Both benchmark runs repeated: the MEMBENCH lines from the second run diff-identical to the first
  for both models.
• make memory-bench is the same command the page documents; the doc's numbers were produced by it,
  not by hand.
• git log --format='%h %G? %s' → G.

Honest scope

• The two models were installed locally for this run (274 MB + 1.2 GB in the machine's Ollama store);
  they are not part of the repository.
• Six paraphrase and four cross-lingual queries: one query moves a column by 17 and 25 points. The page
  says to read the ordering, not the decimals.
• Only Ollama's /api/embed was exercised end to end; the OpenAI-compatible path is still covered by
  httptest only.

Follow-ups

• [memory] Embedding-backed recall, with a rerank stage and a lexical floor #297 S2 stays open: consolidation, quantisation, and the file-level consumer (B4).
• [test] TestCancelledRunMakesNoFurtherRequest fails under the repeated run #295 (the repeated-run flake) remains open and did not reproduce.

Refs #297, #168