MemReader Introduces Active Memory Extraction for AI Agent Systems

Researchers have introduced MemReader, a family of models for active long-term memory extraction in AI agent systems. The work addresses a bottleneck in agent memory: existing systems treat extraction as passive transcription, which produces noisy, low-value entries. The release includes two models. MemReader-0.6B is a compact passive extractor distilled for structured outputs. MemReader-4B is an active extractor trained with Group Relative Policy Optimization (GRPO) that evaluates information value, reference ambiguity, and completeness before deciding whether to write, defer, retrieve context, or discard input. The 4B model operates under a ReAct-style paradigm and achieved state-of-the-art results on knowledge updating, temporal reasoning, and hallucination reduction benchmarks including LOCOMO, LongMemEval, and HaluMem. The authors say MemReader has been integrated into MemOS and is being deployed in real-world applications. The models and API access have been released publicly.