feat: ValorBrain memory provider + BEAM-100K result (71.4% avg score)

[agentxagi]

What

Adds ValorBrain as a memory provider, plus the LLM adapters needed to reproduce our reader/judge pair, and our complete BEAM-100K run.

Result (beam / 100k / rag — 400/400 queries)

Accuracy (avg score per criterion)	71.4%
Binary correct	304/400 (76.0%)
Reader	Gemini 3.6 Flash
Judge	GLM-5.2 (different model family from reader)
Avg retrieve time	5.2s
Avg context tokens	12.6k

The provider

Hybrid memory engine for agents — BM25 + dense vectors (pgvector) + RRF + entity-graph reranking + cross-encoder rerank, on PostgreSQL with row-level security. Retrieval goes through the production path agents use at runtime (/api/v1/memory/prepare), with re-windowed ingest (100k chunks → 8k windows w/ overlap) matching production conversational windows. Plain-HTTP REST, no SDK dependency.

Reproducibility notes

• Full run JSON: outputs/beam/valorbrain/rag/100k.json (answer_llm, judge_llm, per-query scores and judge reasons included)
• Judge override: we pin GLM-5.2 as judge (different family than the Gemini reader) via OMB_JUDGE_LLM — the override mechanism is part of this PR (runner.py), documented in the README
• The score was independently reproduced with a completely different reader (stealth "Ox Alpha" via OpenCode Zen): 70.2% avg score / 76.0% binary on the same 400 queries — evidence the number measures the memory system, not one reader
• Engine: self-hosted; benchmark tenant configuration in the README

Checklist

• Provider registered in the memory registry
• 400/400 queries, no partials
• Judge from a different model family than the reader
• Results JSON + manifest entry included

Happy to adjust to whatever process you prefer for leaderboard inclusion (external_results.json entry, re-run under your infra, etc.).

[vercel]

Someone is attempting to deploy a commit to the Vectorize Team on Vercel.

A member of the Team first needs to authorize it.