docs: matrix v1.2 internal results + #994 FTS-revival impact

[groksrc]

Internal benchmark results record (run 2026-06-12, local, zero API spend).

Headlines

• BM leads QA accuracy on LongMemEval-S (0.617 vs mem0 0.417) and is a close 2nd on ConvoMem (0.792 vs mem0 0.474). mem0 edges retrieval recall but abstains 2-3x as often — its retrieved chunks are less answer-bearing. Retrieval recall ≠ answer quality.
• PR #994 (FTS-revival): full corrected-LoCoMo retrieval +7.9 recall@5 / +10.0 MRR (every category up); QA accuracy +3.7 on the non-adversarial q300 subset.
• multi_hop stays ~0.08 — bottlenecked by BM returning bullet-level chunks that strip document-level context (the session date is in the title), not by FTS. That's the next product fix.
• Full-context is a poor baseline at small-model scale (LongMemEval-S 0.217) but wins on the smaller ConvoMem cs10 (0.825) — confirms full-context only beats retrieval while the corpus fits the model's working window.

Fairness note

mem0 ran in raw-add mode (infer=false) to match the June 10 baseline; its published numbers use infer=true (LLM extraction). A future matrix should run both and document the extraction model.

Run artifacts under benchmarks/runs/ are gitignored; this is the human-readable record.

🤖 Generated with Claude Code