Tools
Amem
By First open-source A-MEM memory plugin for OpenClaw โ Agentic memory with dynamic linking & evolution (NeurIPS 2025)
Install
pip install anthropicConfiguration Example
{
"plugins": {
"allow": ["amem-plugin"],
"entries": {
"amem-plugin": {
"enabled": true,
"config": {
"amemScript": "/Users/you/.amem/amem_client.py",
"condaEnv": "amem",
"condaBase": "/opt/homebrew/Caskroom/miniforge/base",
"userId": "your-username",
"anthropicApiKey": "sk-ant-...",
"dbPath": "/Users/you/.amem/db/"
}
}
},
"slots": {
"memory": "amem-plugin"
}
}
}README
# openclaw-amem
[](LICENSE)
[](https://arxiv.org/abs/2502.12110)
[](https://github.com/openclaw/openclaw)
**A-MEM agentic memory backend for [OpenClaw](https://github.com/openclaw/openclaw)**
An OpenClaw plugin that integrates the [A-MEM](https://arxiv.org/abs/2502.12110) (Agentic Memory) system โ dynamic memory organization with automatic **link generation** and **memory evolution**, backed by ChromaDB + SentenceTransformer + LLM.
> **Note:** This project is an OpenClaw integration of the A-MEM system. For the original research implementation and paper reproduction, see [agiresearch/A-MEM](https://github.com/agiresearch/A-MEM).
---
## Key Features โจ
- ๐ **Dynamic memory organization** based on Zettelkasten principles
- ๐ **Automatic link generation** โ new memories are linked to related existing ones via embedding similarity + LLM judgment
- ๐งฌ **Memory evolution** โ existing memories update their context, tags, and embeddings when new related memories arrive
- ๐ **Hybrid retrieval** โ BM25 + vector search with RRF fusion
- ๐ค **OpenClaw native** โ registers as `memory_search` / `memory_add` tools, works with `plugins.slots.memory`
- ๐ **Local-first** โ ChromaDB + SentenceTransformer, no external vector DB required
---
## What is A-MEM?
A-MEM is a memory system for LLM agents inspired by the Zettelkasten method. Unlike flat vector stores, A-MEM treats memories as a living, self-organizing network:
1. **Note Construction** โ LLM generates keywords, tags, and a context summary on write
2. **Link Generation** โ Top-K candidates (cosine sim > 0.3) are evaluated by LLM; meaningful ones are linked bidirectionally
3. **Memory Evolution** โ Up to 3 linked existing notes have their context/tags/embeddings updated to reflect the new relationship
4. **Hybrid Retrieval** โ BM25 + vector search fused with RRF for robust recall
Paper: _A-MEM: Agentic Memory for LLM Agents_ โ [arXiv:2502.12110](https://arxiv.org/abs/2502.12110)
---
## Architecture
```
OpenClaw Agent
โ
โโโ memory_search(query) โโโบ amem-plugin (OpenClaw plugin)
โโโ memory_add(text) โโโบ โ
โผ
amem_client.py
โ
โโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโ
โผ โผ โผ
ChromaDB SentenceTransformer LLM (Anthropic)
(vector store) (embeddings) (note construction
+ link judgment
+ evolution)
```
---
## Requirements
- Python 3.10+
- conda (miniforge recommended)
- An Anthropic API key (or compatible proxy)
---
## Installation
### 1. Create conda environment
```bash
conda create -n amem python=3.10 -y
conda activate amem
pip install anthropic chromadb sentence-transformers rank-bm25
```
### 2. Place `amem_client.py`
```bash
mkdir -p ~/.amem
cp amem_client.py ~/.amem/
```
### 3. Configure environment variables
| Variable | Default | Description |
|----------|---------|-------------|
| `ANTHROPIC_API_KEY` | *(required)* | Anthropic API key |
| `ANTHROPIC_BASE_URL` | *(Anthropic official)* | Custom base URL (e.g. local proxy) |
| `AMEM_LLM_MODEL` | `claude-opus-4-5` | LLM model for note construction & linking |
| `AMEM_DB_PATH` | `~/.amem/db/` | ChromaDB storage path |
Example:
```bash
export ANTHROPIC_API_KEY=sk-ant-...
export AMEM_DB_PATH=~/.amem/db/
```
### 4. Test
```bash
conda activate amem
python ~/.amem/amem_client.py add "Hello, this is my first memory."
python ~/.amem/amem_client.py search "first memory"
python ~/.amem/amem_client.py list
```
Expected output for `add`:
```
[add] Constructing note...
keywords: ['first memory', 'hello', ...]
tags: ['general']
context: A greeting note marking the first memory entry.
saved note xxxxxxxx-...
[done] Note added: xxxxxxxx-...
```
---
## OpenClaw Plugin Installation
### 1. Copy plugin
```bash
cp -r amem-plugin ~/.openclaw/extensions/
```
### 2. Configure `~/.openclaw/openclaw.json`
```json
{
"plugins": {
"allow": ["amem-plugin"],
"entries": {
"amem-plugin": {
"enabled": true,
"config": {
"amemScript": "/Users/you/.amem/amem_client.py",
"condaEnv": "amem",
"condaBase": "/opt/homebrew/Caskroom/miniforge/base",
"userId": "your-username",
"anthropicApiKey": "sk-ant-...",
"dbPath": "/Users/you/.amem/db/"
}
}
},
"slots": {
"memory": "amem-plugin"
}
}
}
```
### 3. Restart OpenClaw
```bash
openclaw gateway restart
```
### Plugin config reference
| Key | Default | Description |
|-----|---------|-------------|
| `amemScript` | `~/.amem/amem_client.py` | Path to `amem_client.py` |
| `condaEnv` | `amem` | conda environment name |
| `condaBase` | auto-detected | conda base directory |
| `userId` | `default` | User namespace for memories |
| `anthropicApiKey` | `$ANTHROPIC_API_KEY` | API key override |
| `anthropicBaseUrl` | `$ANTHROPIC_BASE_URL` | Base URL override |
| `llmModel` | `$AMEM_LLM_MODEL` | Model override |
| `dbPath` | `$AMEM_DB_PATH` | DB path override |
---
## Usage
Once installed, the agent gains two tools:
### `memory_search`
Search long-term memories using hybrid retrieval (BM25 + vector + RRF).
```
memory_search(query="MetaSmith project status", limit=5)
```
### `memory_add`
Write a new memory. Automatically triggers link generation and memory evolution.
```
memory_add(text="Decided to use ChromaDB for the vector store due to local-first requirements.")
```
---
## How Memory Evolution Works ๐งฌ
When a new memory is added:
1. **Note Construction** โ LLM generates keywords, tags, and a context summary
2. **Embedding** โ SentenceTransformer encodes the enriched note
3. **Link Generation** โ Top-K candidates (cosine sim > 0.3) are evaluated by LLM; meaningful ones are linked bidirectionally
4. **Memory Evolution** โ Up to 3 linked existing notes have their context/tags/embeddings updated to reflect the new relationship
This mirrors the A-MEM paper's core contribution: memories are not static entries but a living, self-organizing network.
---
## Acknowledgements
This project implements the A-MEM architecture proposed in:
> Wujiang Xu et al. _A-MEM: Agentic Memory for LLM Agents_. arXiv:2502.12110, 2025.
Original research repository: [agiresearch/A-MEM](https://github.com/agiresearch/A-MEM)
---
## License
MIT
tools
Comments
Sign in to leave a comment