yuanqianhao/MemSearcher-3B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MemSearcher-3B is a 3.1 billion parameter search agent developed by yuanqianhao, built upon Qwen2.5-3B-Instruct. It utilizes multi-context GRPO, an end-to-end reinforcement learning method, to maintain a compact, iteratively updated memory, ensuring a near-constant context length across multi-turn interactions. This model excels at question answering by efficiently managing memory and performing targeted searches, outperforming ReAct-style agents on various QA benchmarks.

Loading preview...

Overview

MemSearcher-3B is a 3.1 billion parameter search agent, developed by yuanqianhao, based on the Qwen/Qwen2.5-3B-Instruct model. It introduces a novel approach to agentic behavior by employing multi-context GRPO, an end-to-end reinforcement learning method, to manage interaction history. Unlike traditional ReAct-style agents that concatenate entire interaction histories, MemSearcher-3B maintains a compact, iteratively-updated memory, ensuring that the backbone LLM processes only a short question and current memory at each turn. This design keeps the context length nearly constant, significantly reducing computational overhead without sacrificing accuracy.

Key Capabilities

  • Compact Memory Management: Retains only question-relevant information, avoiding long and noisy inputs.
  • Near-Constant Context Length: Achieves stable token counts across multi-turn interactions, optimizing compute and memory usage.
  • Multi-context GRPO: Utilizes an end-to-end reinforcement learning method that propagates trajectory-level advantages for stable optimization of reasoning, search, and memory updates.
  • Strong QA Performance: Demonstrates competitive Exact Match (EM) scores on seven single- and multi-hop QA benchmarks (e.g., 47.0 EM on NQ, 63.8 EM on TriviaQA), outperforming history-concatenation baselines.
  • Integrated Reasoning and Search: The model reasons, decides when to perform a wikipedia_search tool call, and updates its memory with retrieved results, retaining only pertinent information.

Good For

  • Efficient Multi-turn Question Answering: Ideal for applications requiring sustained, accurate QA interactions where context length management is critical.
  • Resource-Constrained Agentic Systems: Suitable for scenarios where minimizing computational resources and memory footprint for LLM agents is a priority.
  • Developing Search-Augmented LLMs: Provides a robust framework for building agents that can effectively reason, search for external information, and manage their internal state.

Finetunes
1 models