Living Brain decouples memory from models: 92.2% LoCoMo, 22.6x fewer tokens
heyshrutimishra · x · 2026-11-04
- Core idea: decouple memory from any single model into one persistent, shared brain — ChatGPT, Claude, Cursor, Codex and your own agents all write to and retrieve from the same memory store.
- Key numbers: 92.2% on the LoCoMo memory benchmark, 464ms median retrieval latency, and up to 22.6x fewer tokens per question versus stuffing full context into the model.
- Pitch: "LLMs have parameters, Living Brain has plasticity" — model weights are frozen, but the memory layer keeps updating and stays portable across agents.
Related event: Living Brain Unveils Cross-Platform Persistent Memory Layer for AI Agents(2 posts)→
More from coding & agent
- Weaviate ships Boost soft re-ranking in v1.38, with an 81k-product search recipe — CShorten30 · 2026-11-04
- AI can write more code, but engineers still must engineer the system — rseroter · 2026-11-04
- Microsoft Clarity adds Agentic Analytics as AI agent traffic nearly doubles to ~2M in 3 weeks — OnlineInference · 2026-11-04
- Agentic Engineering vs AI Engineering: how the two disciplines actually differ — Pavan_Belagatti · 2026-11-04
- Grok Bot Books a Flight End-to-End: Search, Payment, and Bank Verification in One Chat — XFreeze · 2026-11-04
- OpenAI acquires Pierre to make Codex a collaborative AI software-building platform — OpenAINewsroom · 2026-11-04