TrawMem organizes agent memory by task role
Original title: TrawMem: Reframing Agent Memory as Task-Oriented Role-Adaptive Workspaces
On LoCoMo and MemGallery, authors report macro-F1 gains of 2.23–3.14 points and fewer context tokens than compared memory systems.
The paper argues that relevance-based retrieval misses how memories play different roles in a task. TrawMem groups interaction histories into local threads and builds task-oriented workspaces by assigning roles and jointly selecting complementary memories. Experiments on LoCoMo and MemGallery with four open- and closed-source backbones show consistent gains over representative memory systems. On LoCoMo, LLM-Judge scores rise versus SimpleMem while using fewer context tokens.
- Macro-F1 gains are 2.23–3.14 points across the reported comparisons.
- LoCoMo LLM-Judge rises 4.07 points on Qwen3.5-9B versus SimpleMem.
- The same score rises 4.31 points on DeepSeek-V4-Flash.
- Context tokens fall about 12.8% and 5.1% on those two backbones.
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents rely on memory systems to leverage long-term interactions, yet existing systems typically retrieve memories based on their relevance to the current query and use them as contextual evidence.
By Mingyu Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.