ScrapMem: Bio-inspired Framework for On-device LLM Agent Memory
ScrapMem is an innovative framework aimed at facilitating long-term personalized memory management for LLM agents operating on resource-constrained edge devices. It combines various data types into 'Scrapbook Pages' and features Optical Forgetting, a mechanism that gradually lowers the resolution of older memories, achieving storage savings of up to 93% while minimizing less important details. To ensure semantic coherence, it utilizes an Episodic Memory Graph (EM-Graph) that arranges significant events within a causal-temporal framework. In tests conducted on the multimodal ATM-Bench, ScrapMem recorded a leading Joint@10 score of 51.0% and enhanced Recall@10 to 70.3% via structured aggregation. This framework effectively tackles high storage expenses and the intricacies of multimodal data, facilitating efficient on-device memory for AI agents.
Key facts
- ScrapMem is a bio-inspired framework for on-device personalized agent memory.
- It integrates multimodal data into 'Scrapbook Pages'.
- Optical Forgetting reduces resolution of older memories to lower storage cost.
- Episodic Memory Graph (EM-Graph) organizes key events causally and temporally.
- Achieved 51.0% Joint@10 score on ATM-Bench, state-of-the-art.
- Reduces memory usage by up to 93% via optical forgetting.
- Recall@10 improved to 70.3% through structured aggregation.
- Targets resource-limited edge devices for LLM agents.
Entities
Institutions
- arXiv