PlugMem: Why More Memory Makes AI Agents Dumb
Site Owner
发布于 2026-04-18
PlugMem transforms raw interaction history into structured knowledge units, enabling agents to evolve from storing events to actively providing reusable knowledge. Microsoft Research benchmarks show PlugMem outperforms generic retrieval and task-specific memory designs while consuming fewer tokens.
PlugMem: Why More Memory Makes AI Agents Dumb
Here's a counterintuitive conclusion: give an AI agent more memory, and it gets worse at its job.
Microsoft Research proved it. （Source: Microsoft Research Blog "PlugMem: Transforming raw agent interactions into reusable knowledge"）
As interaction logs accumulate, agents have to search through more and more content to find what's relevant. Irrelevant details mix with useful experiences, making retrieval slower and less reliable. The problem isn't storing more experiences—it's organizing them so the agent can quickly find what actually matters right now.
What Cognitive Science Tells Us
PlugMem's design starts from a key distinction in cognitive science.
Memory answers "what happened." But effective decisions rely on knowledge extracted from events—facts and skills.
When someone remembers stumbling through a project last year, that's episodic memory. When they remember "when KPIs are unreasonable, teams走向形式主义," that's knowledge extracted from that event. The former is just context. The latter is decision-relevant.
This distinction applies directly to AI agent memory design. Traditional approaches store conversation logs, document fragments, web sessions—all "events." PlugMem stores facts and reusable skills—the knowledge extracted from those events.
How PlugMem Solves It
PlugMem converts raw interaction history into structured knowledge units. Not retrieving text snippets, but retrieving reusable decision knowledge.
The core logic: raw interactions are standardized into propositional knowledge (facts) and prescriptive knowledge (reusable skills), stored in a knowledge graph rather than a text database. During retrieval, high-level concepts and inferred intents serve as routing signals, pushing the most relevant information to the decision frontier. Before retrieved knowledge enters the agent's context, it's distilled into concise, task-ready guidance—only decision-relevant knowledge consumes the token budget.