80.1 % on Locomo Long-Term Memory Benchmark with a Pure Open-Source Rag Pipeline

I just pushed the current SOTA on the LoCoMo long-term memory benchmark for agents: 80.1 % accuracy using only:

-BGE-large-en-v1.5 (1024d) + FAISS -Custom “MCA” gravitational ranking (keyword coverage + importance + frequency) -BM25 sparse retrieval -Direct Cross-Encoder reranking (bge-reranker-v2-m3) on the full union (~120-150 docs) -Gpt-4o-mini only for final answer generation and judging (everything else is open weights or classic)

Repo: https://github.com/vac-architector/VAC-Memory-System Key tricks that finally broke 80% : -MCA-first filter (coverage ≥ 0.1 → top-30) — catches exact-keyword questions early -Feeding the entire union straight into Cross-Encoder (112–135 documents) instead of pre-filtering -Proper query instruction for BGE-large (the classic “Represent this sentence for searching relevant passages”) The whole pipeline runs in < 3s per query on a single RTX 4090. LoCoMo is currently the hardest public long-term memory benchmark (5.880 real human–agent conversations, multi-hop, temporal, negation, etc.).

Beating Mem0 official baseline by ~12–14 pp with fully open components feels pretty good. Would love feedback, especially from people who are also grinding on agent memory systems.