Tag
This paper introduces HiGram, an evolving hierarchical graph memory framework for LLM agents that features path-level localization and coordinated rewriting to improve retrieval efficiency and answer quality in long-term reasoning tasks.
Proposes Co-E, a training-free system that synchronizes graph and text memory for multi-hop question answering, improving over comparable training-free baselines on six benchmarks.
A Stanford study across 9,842 tasks demonstrates that combining loop engineering with graph memory produces AI agents that outperform standard ones by 38.6%, with 24.1% fewer unnecessary tool calls and 21.7% lower latency.
Discussion of different schools of thought for building memory systems in LLMs, with a focus on graph memory and its potential for human creativity and inductive bias.
A new paper introduces Graph Memory for LLM agents.
The author reflects on the limitations of using flat markdown files for long-term agent memory, which leads to prompt debt as the memory grows, and advocates for graph-based memory representations that retrieve relevant context dynamically.
HeLa-Mem is a bio-inspired memory architecture for LLM agents that models memory as a dynamic graph using Hebbian learning dynamics, featuring episodic and semantic memory stores to improve long-term coherence. Experiments on LoCoMo show superior performance across question categories while using fewer context tokens.