ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical Forgetting

2026-05-05 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

ScrapMem is a new bio-inspired framework designed to provide long-term personalized memory for multimodal Large Language Model (LLM) agents operating on resource-limited edge devices. Developed by Jiale Chang and Yuxiang Ren, ScrapMem addresses challenges like high storage costs and multimodal complexity by integrating data into "Scrapbook Pages." It employs an "Optical Forgetting" mechanism that progressively reduces the resolution of older memories, achieving up to 93% reduction in memory usage. To maintain semantic consistency, ScrapMem utilizes an Episodic Memory Graph (EM-Graph) that organizes key events into a causal-temporal structure. Experiments on the multimodal ATM-Bench demonstrate that ScrapMem achieves a new state-of-the-art Joint@10 score of 51.0% and improves Recall@10 to 70.3% through structured aggregation.

Key takeaway

For research scientists developing LLM agents for edge devices, ScrapMem offers a compelling solution for long-term memory management. You should consider implementing its optical forgetting and EM-Graph mechanisms to significantly reduce storage costs and enhance recall performance, especially when dealing with multimodal data and limited resources. This approach could lead to more efficient and capable on-device agents.

Key insights

ScrapMem enables efficient, personalized long-term memory for LLM agents on edge devices via optical compression and structured graphs.

Principles

• Progressively reduce older memory resolution.
• Organize key events causally and temporally.
• Integrate multimodal data into unified pages.

Method

ScrapMem integrates multimodal data into "Scrapbook Pages," applies "Optical Forgetting" for resolution reduction, and constructs an Episodic Memory Graph (EM-Graph) to maintain semantic consistency through causal-temporal event organization.

In practice

• Reduce memory usage by up to 93% with optical forgetting.
• Improve recall to 70.3% via structured aggregation.
• Achieve 51.0% Joint@10 score on ATM-Bench.

Topics

• ScrapMem Framework
• Optical Forgetting
• LLM Agent Memory
• On-device AI
• Multimodal Data Compression

Code references

• zjunlp/LightMem

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• MemRefine: LLM-Guided Compression for Long-Term Agent Memory — LLM-guided memory compression, prioritizing factual content over surface similarity, effectively manages agent memory within fixed budgets.
• supermemoryai / supermemory — Supermemory provides persistent, context-aware memory for AI, outperforming benchmarks by integrating RAG with dynamic user profiles.
• I Tried Applying JVM Garbage Collection to AI Memory — AI agent memory benefits from intelligent forgetting and lifecycle management, akin to JVM garbage collection.
• ActMem: Bridging the Gap Between Memory Retrieval and Reasoning in LLM Agents — ActMem integrates memory retrieval with active causal reasoning via a structured knowledge graph to enhance LLM agent decision-making.
• Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations — LLM placement within an agent's memory control plane critically determines recovery from various forgetting failure modes.
• SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents — AI agents struggle to discriminate subtle, context-dependent relations (complementary, nuanced, contradictory) among long-term memories.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.