ahmedehabb/Memory-R2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ahmedehabb/Memory-R2 is a 7.6 billion parameter Qwen2.5-7B-Instruct model fine-tuned as a memory-management policy for long-horizon memory-augmented LLM agents. Developed by Sikuan Yan et al., it is specifically designed to decide what to insert, update, or delete in an external memory store during a conversation. This model's unique function is to manage memory, not to generate answers, making it a specialized component for complex agentic systems.

Loading preview...

Overview

Memory-R2 is a specialized 7.6 billion parameter Qwen2.5-7B-Instruct model, developed by Sikuan Yan et al., functioning as a memory-management policy for long-horizon memory-augmented LLM agents. Unlike typical LLMs, its sole purpose is to manage an external memory store by deciding what information to INSERT, UPDATE, or DELETE during a running conversation. It does not answer questions directly; instead, it works in conjunction with a separate answer agent (such as the ahmedehabb/Memory-R2-answer-agent) that reads the memory store it produces.

Key Capabilities

  • Memory Management: Optimizes the content of an external memory store for LLM agents.
  • Long-Horizon Dialogue: Trained specifically for managing memory across extended conversational contexts.
  • Flexible Integration: Designed to be paired with any instruction-tuned LLM as an answer agent.

Training Details

  • Base Model: Qwen2.5-7B-Instruct.
  • Algorithm: Fine-tuned using LoGo-GRPO (turn-level + token-level advantage) with a curriculum of 8, 16, and 32-session rollouts.
  • Dataset: Trained on the LoCoMo long-horizon dialogue dataset.
  • Reward System: Utilizes a per-session cumulative F1 score against gold QA, combined with a memory-compression penalty.

Performance Highlights

When paired with various answer agents, Memory-R2 consistently contributes to strong performance in memory-augmented dialogue systems. For instance, with the SFT+RL trained answer agent, it achieves an F1 score of 51.46 and a BLEU-1 of 44.84.