WujiangXu/A-mem

A‑Mem: Agentic Memory for LLM Agents

What it is – A research implementation of Agentic Memory, a system that lets large‑language‑model (LLM) agents organise their past “memories” dynamically, inspired by the Zettelkasten note‑taking method. The code reproduces the experiments from the paper A‑Mem: Agentic Memory for LLM Agents (NeurIPS 2025).

Key ideas

• When a new piece of information (a memory) is added, the system automatically creates a structured note, tags it, and searches existing memories for related content.
• It then creates explicit links between related memories, forming a graph‑like knowledge network that can evolve as the agent interacts with new data.
• The agent can query this network, retrieve the most relevant memories, and even decide to reorganise or refine them based on the task at hand.

Main features

• Dynamic organisation – memories are linked and re‑indexed on the fly rather than stored in a static vector store.
• Zettelkasten‑style notes – each memory gets a rich, attribute‑filled note (title, description, tags, etc.).
• Agent‑driven management – the LLM itself decides how many memories to retrieve (k), when to create new links, and how to update existing ones.
• Backend‑agnostic evaluation – scripts work with OpenAI’s API, vLLM‑served open‑source models, or Ollama, so you can test the method on any LLM.
• Reproducible experiments – the repository includes the LoCoMo benchmark suite and a k‑sweep script to find the best retrieval size per model.

How to get started

1. Clone the repo and create a Python environment (venv or Conda).
2. Install the required packages with pip install -r requirements.txt.
3. Run the provided evaluation scripts on the LoCoMo dataset:
   • python test_advanced.py – original evaluation (needs OpenAI JSON‑schema support).
   • python test_advanced_robust.py … – robust version that works with OpenAI, vLLM, or Ollama back‑ends. Example for GPT‑4o‑mini:

   python test_advanced_robust.py --backend openai --model gpt-4o-mini \
       --dataset data/locomo10.json --output results.json
   

4. Use bash run_k_sweep.sh to automatically sweep the retrieval‑size (k) hyper‑parameter for each model.

Typical workflow

• Add a memory – the system generates a note, tags it, and searches existing notes for similarity.
• Link creation – if related notes are found, the system adds bidirectional links, enriching the knowledge graph.
• Query – the LLM agent asks a question; the memory system retrieves the top‑k linked notes, which are then fed back to the agent for answer generation.

Intended audience

• Researchers exploring memory architectures for LLM‑based agents.
• Developers who want a reproducible baseline for dynamic, graph‑based memory in their own agents.
• Anyone interested in comparing traditional vector‑store memories with a more structured, link‑driven approach.

Citation

@inproceedings{xu2025amem,
  title={A-Mem: Agentic memory for llm agents},
  author={Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}

License – MIT.

---

This summary is based solely on the repository’s README; no external information has been added.