MemoryArena

Minimal, modular framework for connecting agents, environments, and memory.

Implementation for MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (https://arxiv.org/abs/2602.16313).

This code is preview version. We are still actively maintaining and improving this codebase.

Quick Start

[Important] Check setup_web_shopping.md, setup_travel.md, setup_web_search_env.md, and setup_formal_reasoning.mdto follow step-by-step structions for each test environment.

Repository Structure

• agent/: task agent implementations.
• env/: environment server, client, and environment systems.
• memory/: memory client and memory systems including long-context, letta, mirix, mem0, mem0-g, ReasoningBank, BM25, Text-embedding RAG, GraphRAG, and MemoRAG.

API Keys

• Make sure you have your OPENAI_API_KEY, OPENAI_BASE_URL, GOOGLE_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY, OPENROUTER_API_BASE_URL set ready in either your bashrc file or in configs following each setup md.
• For Letta, Mirix, Mem0 (including Mem0-g), make sure you have their memory system api keys ready LETTA_API_KEY, MIRIX_API_KEY, MEM0_API_KEY in your bashrc file.

Example Flow

1. Task prompt → memory wraps prompt
2. Agent generates action
3. Env step() executes tool or accepts final
4. Observation + reward returned
5. Memory stores action/observation/reward(optional)

Cite Our Paper:

If you are using this repo, please cite our paper at:

@article{he2026memoryarena,
  title={MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks},
  author={He, Zexue and Wang, Yu and Zhi, Churan and Hu, Yuanzhe and Chen, Tzu-Ping and Yin, Lang and Chen, Ze and Wu, Tong Arthur and Ouyang, Siru and Wang, Zihan and others},
  journal={arXiv preprint arXiv:2602.16313},
  year={2026}
}