Agent memory stops ranking by similarity alone: TrawMem assigns task roles, lifting LoCoMo judge scores by about 4 points

Agent memory is now selected by the role a memory plays in a task rather than by relevance to the current query, so the same memory can serve different roles across tasks. Existing memory systems retrieve by relevance to the current query and use the results as contextual evidence; relevance alone does not describe how a memory contributes to solving the task. The change is tested on LoCoMo and MemGallery with four open- and closed-source backbones, where macro-F1 improves by 2.23 to 3.14 points over representative memory systems.

The system first splits interaction histories into local threads, then assigns roles to memories and jointly selects complementary ones into a task-oriented workspace. The approach, TrawMem, treats memory as a task-oriented role-adaptive workspace, selecting and organizing memories by task roles and relations instead of ranking them by semantic similarity alone. On LoCoMo, LLM-Judge scores exceed SimpleMem by 4.07 and 4.31 percentage points on Qwen3.5-9B and DeepSeek-V4-Flash, while requiring roughly 12.8% and 5.1% fewer context tokens.

This report rests on summary-level material covering evaluations on LoCoMo and MemGallery with four open- and closed-source backbones, plus the LoCoMo comparison against SimpleMem on LLM-Judge score and context tokens.

The summary reports only the macro-F1 gain range and the LoCoMo LLM-Judge and token comparison; per-backbone results and the specific criteria for role assignment are not detailed.