Introduction

amem is an agentic-memory stack for LLM agents — memory that constructs, links, and evolves notes like a Zettelkasten instead of dumping flat vector rows. It is an open-source implementation of the A-MEM research (NeurIPS 2025), written in TypeScript on top of Qdrant + local Transformers.js. No Python required.

The amem stack

amem is a monorepo you can adopt one piece at a time:

Package	Role	Status
@heichaowo/amem-core	Engine — note construction, evolution, hybrid retrieval. Framework-agnostic.	shipping
openclaw-amem	OpenClaw Plugin — drops A-MEM into OpenClaw's memory slot.	shipping
amem-api	Server — single-writer HTTP + MCP service so many processes share one store.	coming soon

New here? Start with the OpenClaw Plugin →. It's the fastest way to give an agent evolving long-term memory today — the amem-core engine is bundled inside it, so there's nothing extra to install.

---

What is A-MEM?

A-MEM is a memory architecture for LLM agents inspired by the Zettelkasten method. Unlike a flat vector database, A-MEM maintains memory as a living, self-evolving semantic graph. This is the behavior the amem-core engine implements — every consumer in the stack inherits it.

The memory lifecycle

1. Note Construction — On write, the LLM extracts keywords, tags, a context summary, and categorizes the note (Technical, Business, Personal, Project, Research, System, General).

2. Link Generation — Retrieves top-6 candidates; the LLM judges whether to link bidirectionally (similarity > 0.3).

3. Memory Evolution & Strengthening — Up to 3 linked memories have their attributes evolved based on the new context, potentially triggering additional links.

4. Hybrid Retrieval — Fuses vector search (local ONNX multilingual-e5-small, 384-dim) and BM25 using Reciprocal Rank Fusion (RRF), boosted by retrieval frequency (heat).

5. 2-hop BFS Graph Expansion — After RRF top-K selection, BFS traverses the link graph up to 2 hops, appending up to 8 contextually linked notes. Each candidate passes an embedding relevance gate (cos-sim ≥ 0.25) before admission.

6. Per-Agent Memory Isolation — Each agent operates in its own private namespace (agent_id filter in Qdrant). Memories written by main are invisible to dev by default. A shared scope (explicit agent_id="shared") allows publishing to all agents. See Agent Isolation for full details.

---

Architecture

How the OpenClaw Plugin wires the engine in-process:

OpenClaw Agent
     │
     ├── memory_search(query)  ──►  openclaw-amem plugin (TypeScript, in-process)
     └── memory_add(text)      ──►       │
                                         ▼
                          ┌──────────────┼──────────────┐
                          ▼              ▼               ▼
                       Qdrant     Transformers.js  LLM (Anthropic/OpenAI)
                    (vector store)  (ONNX embed)   (CRUD decision
                      :6333        384-dim local    + link judgment
                   agent_id ISO   + Jieba BM25     + evolution)

When amem-api ships, this same engine runs as a shared single-writer service instead of in-process — so multiple agents and processes read and write one store.

---

Academic Background

Based on the paper: A-MEM: Agentic Memory for LLM Agents — arXiv:2502.12110 (NeurIPS 2025). For the original research implementation, see agiresearch/A-MEM.

@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 (NeurIPS)},
  year={2025}
}