A post about a paper describes agents that write, search and organize notes, learning when memory is useful. According to the post, this improved an open 32B model's performance by 2x to 4x in long games.
Published on July 6, 2026, the post describes AutoMem, a paper about agents that write, search and organize notes, and learn when to use memory. According to the post, the approach improved an open 32B model's performance by 2x to 4x in long games.
The idea is for agents to manage their own memory to help with long-horizon tasks without expanding the context window. To verify the scope and results, consult the post and the original paper; the figures above are those reported by the post, not an independent evaluation.