Researchers at City University of Hong Kong and Tencent built SoMem, a system that lets AI agents keep their own wiki of past chats, and it beat the best rival memory systems on three tests.

Key Takeaways

  • SoMem beat the strongest rival on LoCoMo, BEAM-100K and LongMemEval-S by 2.66, 3.10 and 2.00 percentage points
  • The agent decides what to summarize, how to group notes and when to update a topic
  • Every summary stays linked to the original chats, so the agent can check its own notes
  • The memory can be reused by other models, and the code is public

The team posted the paper on arXiv on Nov. 9. Most AI agents cannot learn after they are deployed because their model weights are fixed, so anything new has to be stored in outside memory.

Existing systems such as Mem0 and MemoryBank mostly save compact records under fixed filing rules. The authors say those rules cannot easily keep up as tasks and conversations change.

Also Read: RoboScript Teaches Robots A Task From One Demo, Beating Trained Rivals By 20 Points

How SoMem Works

In SoMem, memory is a set of linked documents, like a wiki the agent edits itself. The agent writes topic pages from records of past chats and revises them when new information arrives, such as a meeting moving to a new venue.

The system checks each edit and builds three extra views from the notes, a timeline, a map of connections and a list of recent changes. When a question comes in, the agent searches, follows links and can open the original chat to check a summary, much like checking a footnote.

Memory test SoMem lead over best rival
LoCoMo +2.66 points
BEAM-100K +3.10 points
LongMemEval-S +2.00 points

Small Gains, But Reusable

The margins are a few percentage points and come from the authors' own experiments in a paper that has not been peer reviewed. The authors say the memory one model builds can be read by another, and that letting the model reason longer improves its answers further.

They call the result a feasibility finding. Agents can organize and use their own long-term memory with ordinary documents instead of a fixed filing scheme, they write, and the code is on GitHub.

Read Next: Nvidia’s Tiny 3D AI Matches Models 10 Times Its Size By Reusing One Layer

Originally published on Fathom News.