Add AutoMem memory provider + neutral-run results

[jack-arturo]

Summary

Adds AutoMem — an MIT-licensed graph + vector memory service (FalkorDB + Qdrant behind a Flask API) — as an AMB provider, plus committed results from a neutral AutoMem 0.16.0 run.

The provider is self-spinning: initialize() brings the whole stack up via docker compose, waits for /health, and an atexit-registered cleanup() tears it down (docker compose down -v). A maintainer can reproduce every number below with only a GEMINI_API_KEY — embeddings are local via FastEmbed, so no embedding API keys are needed.

What's included

• src/memory_bench/memory/automem.py — provider. ingest() batches documents to POST /memory/batch (chunked on sentence/paragraph boundaries, timestamps backdated to the source), then waits for enrichment to settle; retrieve() queries GET /recall with graph-relation expansion.
• Registry entry in src/memory_bench/memory/__init__.py and a display entry in catalog.json.
• automem_compose.yml — pinned stack: ghcr.io/verygoodplugins/automem:amb-v1, falkordb/falkordb:v4.18.3, qdrant/qdrant:v1.11.3.
• AUTOMEM_REPRODUCE.md — one command per split.
• tests/test_automem_provider.py — HTTP-contract + chunking unit tests (no Docker); 10 passing.
• outputs/**/automem/rag/*.json.gz — committed run results (one canonical run per split).

Results — AutoMem 0.16.0, rag mode

Answerer gemini-3.1-pro-preview, judge gemini-2.5-flash-lite, embeddings FastEmbed BAAI/bge-base-en-v1.5 (768d).

Dataset / split	Accuracy
beam / 100k	67.6%
beam / 500k	65.6%
beam / 1m	63.8%
beam / 10m	57.4%
locomo / locomo10	85.1%
longmemeval / s	74.4%
personamem / 32k	76.1%

AutoMem's strength is long-context scaling and efficiency: accuracy holds across the BEAM 100k → 10M sweep while the mean context fed to the answerer stays ~2.6–4.8k tokens at every scale (the committed output files record avg_context_tokens per run). Verbatim conversational recall (locomo / longmemeval / personamem) trails the current leaders — reported here as-is, not cherry-picked.

Reproduce

GEMINI_API_KEY=... \
OMB_ANSWER_LLM=gemini OMB_ANSWER_MODEL=gemini-3.1-pro-preview \
OMB_JUDGE_LLM=gemini  OMB_JUDGE_MODEL=gemini-2.5-flash-lite \
uv run omb run --memory automem --dataset locomo --split locomo10

Swap --dataset/--split for the others (longmemeval/s, personamem/32k, beam/{100k,500k,1m,10m}); the env block is identical. Ports are chosen per run, so repeat/concurrent runs don't collide, and a crash still tears the stack down. Full details in AUTOMEM_REPRODUCE.md.

The committed outputs are self-run; a maintainer re-run against the pinned image is welcome to confirm them on the neutral board.

[vercel]

@jack-arturo is attempting to deploy a commit to the Vectorize Team on Vercel.

A member of the Team first needs to authorize it.