Built a memory system solo in 16 days that beats every funded AI memory company on LongMemEval (96.2%, open source)

Just published agentmemory V4 — scored a world record 96.2% on LongMemEval, the standard benchmark for long-term AI agent memory. This beats: \- PwC Chronos: 95.6% \- Mastra: 94.87% \- OMEGA: 93.2% (raw) \- Supermemory: 85.86% \- Emergence AI: 86% \- Zep: 71.2% Built it alone. Mid-range gaming PC (i3-12100F). $1,000 total. 16 days. Uses Claude Opus as generator and GPT-4o as judge but the retrieval architecture is the interesting part — HNSW + BM25 + cross-encoder + knowledge graph + temporal grounding, all in a single SQLite-backed system. Open source (MIT): [github.com/JordanMcCann/agentmemory](http://github.com/JordanMcCann/agentmemory)