AlekseiMarchenko/central-intelligence

New benchmark results and documentation parity. Central Intelligence's v2 hybrid retrieval pipeline (pgvector HNSW + BM25 + RRF fusion + bge-reranker-v2-m3 cross-encoder) scores among the top results published on both LifeBench and LongMemEval.

Benchmark Results

LongMemEval (ICLR 2025) — 75.0%

Conversational memory across 500 questions: single-session recall, multi-session reasoning, temporal reasoning, knowledge updates, preference tracking.

| Overall | Single-session | Multi-session | Temporal | Preference |
|---------|----------------|---------------|----------|------------|
| 75.0% | 91.9% | 66.2% | 69.9% | 76.7% |

Answer model: gpt-5.4-mini. Judge: gpt-4o.

LifeBench (2026) — 52.2%

Long-term multi-source memory: 2,003 questions across 10 users, 51K real-world events (messages, calendar, health records, notes, calls). The hardest published memory benchmark.

| Overall | Info Extraction | Multi-hop | Temporal | Nondeclarative |
|---------|-----------------|-----------|----------|----------------|
| 52.2% | 47.2% | 52.9% | 46.4% | 64.1% |

Answer model: gpt-5.4-mini. Judge: gpt-4.1-mini.

Evaluation harness: lifebench-eval.

What Got Us Here (since v1.2.1)

The journey from v1.2.1 to v1.2.2, measured on LongMemEval:

| Change | Delta |
|--------|-------|
| Reranker upgrade: MiniLM → bge-reranker-v2-m3 + top-K=10 | ~+18pp |
| Answer model: gpt-4o-mini → gpt-5.4-mini | ~+5pp |
| Answer prompt: CoT counting + factual precision + multi-hop grounding | ~+3pp |
| Infrastructure fixes (ef_search, performance CPU, 4GB RAM) | enabler |

Attribution is approximate — each step was measured through the benchmark's run-to-run variance. The architecture that reached 75% is the clean v2 pipeline: pgvector HNSW + BM25 with RRF fusion, temporal decay, and a cross-encoder reranker. No fact extraction or graph traversal during recall.

Changes Since v1.2.1

Retrieval

• Upgraded reranker to bge-reranker-v2-m3 (8K context, +14% nDCG@10 over MiniLM)
• top-K=10 retrieval (NeurIPS 2024 finding: accuracy saturates at k=10, degrades with more docs)
• Set hnsw.ef_search = 400 globally (was silently capped at ~40 results)
• Reverted to v2 architecture (vector + BM25 + reranker) — removed query decomposition and fact-path that regressed multi-session scores

Benchmark infrastructure

• Added --answer-model parameter to benchmark entrypoint
• Performance CPU on benchmark VM (fixes extraction stalls under shared-CPU throttling)
• DB-driven extraction queue (removed in-memory queue that caused zombie processes)

Documentation

• README Benchmarks section now includes full category tables for both LifeBench and LongMemEval
• Landing page at centralintelligence.online matches README numbers exactly
• Answer/judge model attribution documented inline for reproducibility

Try It

npx central-intelligence-local signup

Full docs: README · Benchmarks: live results