Benchmarks

July 15, 2026 · View on GitHub

Run from repo root unless noted. Install deps:

pip install chromadb pytrec-eval-terrier "fluctlightdb[native]>=0.5.2"
# or dev: pip install -e sdks/python && ./scripts/install-native.sh

# LoCoMo one-command reproduce (downloads data, Fabric on, checks frozen cert)
make reproduce-locomo
# Frozen baseline: benchmarks/results/locomo-invented-stack-engine-2026-07-13.json
# LoCoMo @150: 96.8% MiniLM-384 / 97.0% mpnet-768 (native Rust, no expansion; @5=72.6%/75.1%).
# The historical 99.0% was ±3 neighbor-expansion inflation, not the engine — deprecated. Evidence-recall ≠ QA.
ScriptPurposeData
beir_bench.pyCertified IR (nDCG@10, Recall@10/100)BEIR SciFact
agent_memory_bench.pyAgent-specific: paraphrase, provenance, persistenceBuilt-in
locomo_bench.pyLoCoMo long-dialogue evidence recallLoCoMo
longmemeval_bench.pyLongMemEval session recall (--mode brain default)LongMemEval
longmemeval_e2e.pyLongMemEval E2E (brain retrieve → reader → judge)Same + API key
brain_memory.pyBrain-native ingest (CLS sleep, fact engrams, completion)(library)

LongMemEval modes

ModeAPIUse
brain (default)connect_brain()Full agent path: dentate separation, graph spread, CLS sleep, cortex boost, fact/turn engrams, pattern completion
convconnect_conv()Hybrid RAG (LoCoMo-style), no sleep
indexconnect_index()IR-only vector-fast (Chroma-class); not the agent brain — legacy retrieval baseline

E2E profiles: brain (CLS + Chain-of-Note + GPT-5), max (GPT-5 + CoT + top-50), standard (gpt-4o).

E2E full 500: bash scripts/run-longmemeval-e2e-500.sh — see docs/LONGMEMEVAL_E2E.md.

Paper citations and protocol: docs/BENCHMARKS.md.

# BEIR
BEIR_DATA=/tmp/beir BEIR_DS=scifact MODE=index PYTHONPATH=sdks/python python benchmarks/beir_bench.py

# FAMB
PYTHONPATH=sdks/python python benchmarks/agent_memory_bench.py --mode agent

# LoCoMo / LongMemEval (after data download)
LOCOMO_DATA=/tmp/locomo PYTHONPATH=sdks/python python benchmarks/locomo_bench.py
LONGMEMEVAL_DATA=/tmp/LongMemEval/data PYTHONPATH=sdks/python python benchmarks/longmemeval_bench.py