AI News
All stories
Official

SoMem lets agents organize their own memory as linked documents

City University of Hong Kong and Tencent Jarvis Lab propose self-organized memory, in which agents decide how to summarize and link their own notes.

Research · · 1 min read

City University of Hong Kong and Tencent Jarvis Lab propose SoMem, an agent memory design. The model decides how its own notes are structured, rather than writing them into a fixed schema of record types and grouping rules. The authors argue that fixed schemas are hard to adapt as tasks and interactions change. The arXiv abstract frames the question as whether LLM agents can organize their own memory using ordinary documents. Their answer is a framework that stores memory as linked documents.

The agent chooses what to summarize, how to group related information into topics, and when to revise existing topics. Each summary it writes keeps a link back to the original interactions, which are archived verbatim. The agent can read that archive but cannot edit it. A runtime validates each proposed edit before installing it, then derives timeline, relation graph and recent-memory views from the organized content.

Retrieval works the same way in reverse. At question time, the agent decides what to search and read, follows links between documents, and opens the original evidence when it needs to confirm an answer. This is the agent-controlled reading the write-up describes.

SoMem beats the strongest baseline by 2.66 percentage points on LoCoMo, 3.10 on BEAM-100K and 2.00 on LongMemEval-S. The DAIR.AI write-up gives these as points, while the arXiv abstract calls them percentage points. The paper's abstract also says the constructed memory can be reused across models, and that additional reasoning effort can further improve answer quality.

DAIR.AI's practical pitch is that an agent already reading and writing files could try linked notes with source evidence before adding a vector store. The paper shows something narrower. An agent that writes its own linked documents, keeps raw evidence immutable and retrieves by following links can outperform the baselines it was tested against. The tests covered three long-conversation memory benchmarks.

What changed

SoMem lets agents organize their own memory as linked documents instead of writing it into fixed schemas of record types and grouping rules.

Type: Research · Applies to: LoCoMo, BEAM-100K, LongMemEval-S

Why it matters

Teams building agents with memory could avoid fixed schemas, and memory written by one model was reported to work when read by another.

What you can do

Consider linked notes with source evidence as a simple memory design to test before adding a vector store.

Original source

Where we saw it

1 source

DAIR.AI (Elvis Saravia)

City University of Hong KongTencent Jarvis Lab

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works