TrawMem
A framework that redefines agent memory systems, arguing that memory should be selected and organized by its role and relationships within a task rather than retrieved solely by semantic relevance to a query. It implements this through local threads and a joint selection mechanism.
Timeline (last 90 days)
TrawMem proposes the "task-oriented role-adaptive workspace" framework, arguing that memories should be selected and organized by their roles and relationships in the task, rather than retrieved solely by relevance.
TrawMem changes the unit of memory selection from individual snippets to role combinations, and changes the organization of memories from a flat list to a structured workspace.
TrawMem organizes interaction history into local threads, i.e., cohesive segments split by interaction continuity.
TrawMem builds a task-oriented workspace, assigns roles to memories, and performs joint selection of complementary memories.
The paper evaluates on two benchmarks, LoCoMo and MemGallery, using four open-source and closed-source backbone models, improving macro-F1 by 2.23 to 3.14 points.
On the LoCoMo benchmark, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 percentage points on Qwen3.5-9B and 4.31 percentage points on DeepSeek-V4-Flash.
On the LoCoMo benchmark, TrawMem reduces context token consumption relative to SimpleMem by about 12.8% (Qwen3.5-9B) and 5.1% (DeepSeek-V4-Flash).
The current summary covers only two benchmarks and four backbone models, and has not yet addressed more complex agent scenarios such as longer interaction histories and multi-turn tool calls.
TrawMem is a structural reworking of the memory system rather than adding a reranking module on top of a retriever; it changes the granularity and basis by which memories are selected and organized.
All Facts (9)
TrawMem proposes the "task-oriented role-adaptive workspace" framework, arguing that memories should be selected and organized by their roles and relationships in the task, rather than retrieved solely by relevance.
50%UnverifiedTrawMem changes the unit of memory selection from individual snippets to role combinations, and changes the organization of memories from a flat list to a structured workspace.
50%UnverifiedTrawMem organizes interaction history into local threads, i.e., cohesive segments split by interaction continuity.
50%UnverifiedTrawMem builds a task-oriented workspace, assigns roles to memories, and performs joint selection of complementary memories.
50%UnverifiedThe paper evaluates on two benchmarks, LoCoMo and MemGallery, using four open-source and closed-source backbone models, improving macro-F1 by 2.23 to 3.14 points.
50%UnverifiedOn the LoCoMo benchmark, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 percentage points on Qwen3.5-9B and 4.31 percentage points on DeepSeek-V4-Flash.
50%UnverifiedOn the LoCoMo benchmark, TrawMem reduces context token consumption relative to SimpleMem by about 12.8% (Qwen3.5-9B) and 5.1% (DeepSeek-V4-Flash).
50%UnverifiedThe current summary covers only two benchmarks and four backbone models, and has not yet addressed more complex agent scenarios such as longer interaction histories and multi-turn tool calls.
50%UnverifiedTrawMem is a structural reworking of the memory system rather than adding a reranking module on top of a retriever; it changes the granularity and basis by which memories are selected and organized.
50%