Tutorials on Plugmem

Traditional memory systems for LLM agents face critical limitations that hinder performance and scalability. Research shows that 72% of AI agents struggle to effectively reuse long interaction histories due to raw memory logs being noisy, verbose, and contextually irrelevant. For example, unstructured memory retrieval often overwhelms agents with redundant data, leading to higher token costs and lower decision accuracy. In benchmarks like LongMemEval, agents using raw memory achieved only 71.2% accuracy in multi-turn dialogue tasks, while structured memory systems like PlugMem improved this to 75.1% using 362.6 memory tokens per sample -a 20% efficiency boost. This highlights the urgent need for knowledge-centric memory that prioritizes semantic and procedural knowledge over raw experience. As shown in the Benchmarking and Evaluating PlugMem section, these improvements are validated through rigorous performance metrics. PlugMem redefines memory design by organizing interactions into a graph-based knowledge system that separates propositional facts (semantic knowledge) and prescriptive strategies (procedural knowledge). This approach addresses two major pain points: For example, in HotpotQA (a multi-hop question-answering benchmark), PlugMem achieved 61.4% accuracy by linking semantic concepts like “Jim Croce” to his birth year through a two-step reasoning process, whereas traditional systems scored 57.8% using 2–3× more tokens. This efficiency stems from hierarchical retrieval that prioritizes high-level concepts before diving into episodic details. The Encoding Propositional Knowledge in the Memory Graph section details how this structured approach enables precise knowledge extraction.