PlugMem: Turning Raw Agent Interactions into Reusable Knowledge

The longer AI agents accumulate interaction history, the harder it becomes to use. Irrelevant content gets mixed in, driving up retrieval costs, and it becomes difficult to pull out exactly the right information at the moment it's needed.

PlugMem addresses this problem by not storing raw interactions as-is, but instead converting them into structured knowledge organized as facts and skills. It transforms records such as conversations, documents, and web sessions into propositional knowledge and prescriptive knowledge, accumulates them in a memory graph, and makes it possible to select and reuse only the knowledge relevant to the current task.

The architecture consists of three stages.

• Structure: Standardizes raw interactions and converts them into facts and reusable skills.
• Retrieval: Retrieves knowledge units suited to the current task, rather than long text fragments.
• Reasoning: Compresses the retrieved knowledge into short, actionable instructions and inserts them into the agent's context.

The authors explain that this approach starts from a distinction in cognitive science: remembering events is different from knowing facts and skills. The purpose of memory, they argue, is not merely to store the past, but to convert it into knowledge needed for the next action.

Performance was evaluated across 3 benchmarks with different memory requirements.

• Answering questions in long multi-turn conversations
• Finding facts across multiple Wikipedia documents
• Making decisions while navigating the web

As a result, PlugMem consistently outperformed general-purpose retrieval methods and task-specific memory designs, while simultaneously using a smaller memory token budget. The authors argue that what matters more than the sheer size of memory is how usefully the needed information is delivered at the right time.

To this end, they also proposed a metric that looks at utility together with context consumption, rather than the sheer amount of information. Under this criterion as well, PlugMem provided more decision-relevant information than other methods while using less context, demonstrating that converting experience into knowledge is more efficient than storing and retrieving raw logs as-is.

PlugMem is not a memory module tailored to a single specific task. It is designed as a general-purpose memory layer that can be attached to a variety of settings, including conversational agents, knowledge retrieval, and web agents, and can achieve additional performance gains when combined with task-specific techniques.

The code and experimental results are publicly available on GitHub, enabling reproduction and follow-up research.