SoMem paper: LLM agents organize their own memory via linked documents, beating baselines
omarsar0 · x · 2026-11-12
A CityU Hong Kong + Tencent Jarvis Lab paper introduces SoMem, a self-organized memory framework where the agent decides what to summarize, which topics to create, and when to revise — stored as ordinary linked documents, with every summary linked back to original interactions as evidence and runtime-derived temporal, relational, and recency views. It outperforms the strongest baselines on LoCoMo, BEAM, and LongMemEval-S by 2.66, 3.10, and 2.00 points respectively; the constructed memory transfers across models, and extra reasoning effort further improves answers. Code is open source.
More from coding & agent
- Mental models, not coding agents, are now the biggest barrier to great products — mrjonfinger · 2026-11-12
- A one-line prompt that makes your coding agent spell out its assumptions before acting — SIGKITTEN · 2026-11-12
- In the AI era ideas ship in an evening — moats are user interviews, not code — Al_Grigor · 2026-11-12
- Peter Yang's 500-hour guide: when to use Claude Fable 5.2 vs Opus 5.5 — petergyang · 2026-11-12
- Which Claude model for which job: a 5-scenario real-work comparison — petergyang · 2026-11-12
- SGLang ported to IBM Spyre in under a week via AI coding agents, 3.6x throughput — PyTorch · 2026-11-12