yuanqianhao/MemSearcher-7B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

MemSearcher-7B by yuanqianhao is a 7.6 billion parameter search agent built on Qwen2.5-7B-Instruct, designed for efficient question answering. It utilizes multi-context GRPO for end-to-end reinforcement learning, maintaining a compact, iteratively-updated memory to keep context length nearly constant across turns. This model excels at multi-hop QA benchmarks by reasoning, searching, and managing memory effectively without concatenating entire interaction histories.

Loading preview...

Overview

MemSearcher-7B is a 7.6 billion parameter search agent developed by yuanqianhao, based on the Qwen/Qwen2.5-7B-Instruct architecture. Its core innovation lies in its memory management system, which significantly differs from traditional ReAct-style agents. Instead of concatenating the entire interaction history, MemSearcher-7B maintains a compact, iteratively-updated memory, ensuring that the backbone LLM processes only the current question and a concise memory. This approach results in a nearly constant context length across multi-turn interactions, drastically reducing computational overhead and memory requirements without compromising accuracy.

Key Capabilities

  • Efficient Memory Management: Keeps only question-relevant information in memory, avoiding long and noisy inputs.
  • Stable Context Length: Achieves near-constant token count across multiple turns, leading to reduced compute and memory usage.
  • End-to-End Reinforcement Learning: Trained with multi-context GRPO, which propagates trajectory-level advantages to every turn for stable optimization of reasoning, search, and memory updates.
  • Strong QA Performance: Demonstrates competitive performance on seven single- and multi-hop QA benchmarks, outperforming history-concatenation baselines.

How It Works

At each turn, the model processes only the (question, memory_{i-1}). It then:

  1. Reasons within <think>...</think> tags.
  2. Searches using a wikipedia_search tool call if knowledge is insufficient, or provides an answer in \boxed{} if the memory is sufficient.
  3. Updates the memory by retaining only information from retrieved results that is relevant to the question.

Good For

  • Question Answering Systems: Particularly effective for complex, multi-hop QA tasks where efficient information retrieval and synthesis are crucial.
  • Resource-Constrained Environments: Its ability to maintain a compact context makes it suitable for applications where minimizing compute and memory usage is important.
  • Agentic Workflows: Designed as a search agent, it's ideal for scenarios requiring dynamic reasoning, tool use (like Wikipedia search), and intelligent memory management.

Finetunes
1 models