amem

Appearance

amemAgentic memory that evolves

An open-source agentic-memory stack for LLM agents — memories evolve, not just accumulate. Graph linking, hybrid retrieval, LLM-driven evolution. TypeScript, local embeddings, no Python.

Get Started
View on GitHub
npm
A-MEM Logo
🔄

Dynamic Memory Network

Inspired by Zettelkasten. Memories are stored as nodes in a graph with automatic bidirectional link generation and LLM-verified connections — not flat vector rows.

🧬

Memory Evolution & Consolidation

Linked memories evolve context, tags, and embeddings when new details arrive. Same-day semantic duplicates (≥ 0.80 similarity) are auto-merged, and daily consolidation at 02:30 keeps the graph clean.

🚦

LLM CRUD Gate & Quality Controls

Hooked into agent_end to run NEW / UPDATE / DELETE / NONE decisions on every conversation. A write-time quality gate rejects low-quality content; memory_quality_scan surfaces stale or conflicting entries.

🔍

Hybrid Retrieval & Heat Tracking

BM25 + dense vector hybrid (RRF) with 2-hop BFS graph expansion. Frequently retrieved memories get a logarithmic heat boost dampened by time decay, so fresh facts stay on top.

🧠

Knowledge vs Episodic

Notes are auto-classified as memory (episodic) or knowledge (durable). Knowledge notes carry topic tags and skip consolidation merging and time-decay penalties.

🀄

Chinese-Optimized BM25

Uses Jieba (via @node-rs/jieba) for CJK word segmentation, dramatically improving recall for Chinese queries. English and mixed text fall back to whitespace tokenization automatically.

🔐

Per-Agent Memory Isolation

Each agent operates in its own private namespace — memories written by main are invisible to dev by default. An explicit shared scope plus owner/readers/writers fields on every note control cross-agent access.

Install the OpenClaw plugin ​

The fastest way to try amem today is the openclaw-amem plugin — the amem-core engine is bundled inside, so there is nothing else to install.

bash
openclaw plugins install clawhub:openclaw-amem
bash
openclaw plugins install openclaw-amem

Point OpenClaw's memory slot at openclaw-amem and your agent remembers across sessions — linking, evolving, and consolidating on its own. → Full installation guide

The amem stack ​

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

PackageRoleStatus
@heichaowo/amem-coreEngine — note construction, evolution, hybrid retrieval. Framework-agnostic.shipping · npm · also bundled in the plugin
openclaw-amemOpenClaw Plugin — drops A-MEM into OpenClaw's memory slot.shipping · npm · ClawHub
amem-apiServer — single-writer HTTP + MCP service so many processes share one store.coming soon

Grounded in research ​

openclaw-amem implements A-MEM: Agentic Memory (NeurIPS 2025) — memory that constructs, links, and evolves notes like a Zettelkasten instead of dumping flat vector rows. Embeddings run locally via Transformers.js (384-dim) and are stored in Qdrant. No Python, no external embedding API. MIT-licensed.

In a hurry? Jump straight to the Quick Start.

Released under the MIT License.

Based on A-MEM (NeurIPS 2025)