ClawMem

Local-first memory engine for OpenClaw — durable, auditable, reversible.

"Markdown stays the source of truth, but recall is persistent, searchable, and graph-aware."

---

What is ClawMem?

ClawMem is a standalone, open-source alternative to SuperMemory — locally hostable, zero-server, and designed as a native memory-slot plugin for OpenClaw.

Why ClawMem over Supermemory or Mem0?

Feature	Supermemory	Mem0 OSS	ClawMem
Graph memory (UPDATE/EXTEND/DERIVE)	✅ Neo4j required	✅ Neo4j required	✅ Kùzu embedded (zero ops)
Vector search	✅	✅	✅ sqlite-vec
Hybrid search (vector + keyword)	✅	❌	✅ FTS5 + vector
User profiles (static + dynamic)	✅	❌	✅
Contradiction resolution	✅	Partial	✅ Graph UPDATE chains
Temporal reasoning	✅	❌	✅
Local-first (no cloud required)	❌	✅	✅
Zero-server (fully embedded)	❌	❌	✅
Markdown two-lane sync	❌	❌	✅
Reversible (full audit log)	❌	❌	✅
OpenClaw memory-slot plugin	✅	✅	✅ Native

---

Quick Start

As an OpenClaw plugin (30 seconds)

openclaw plugins install @clawmem/openclaw

Add to your OpenClaw config:

{
  "plugins": {
    "slots": {
      "memory": "@clawmem/openclaw"
    }
  }
}

Configure ClawMem (point to your local LLM + embedder):

{
  "clawmem": {
    "llm": { "baseURL": "http://127.0.0.1:8080/v1", "model": "deepseek-r1" },
    "embedder": { "baseURL": "http://127.0.0.1:8082/v1", "model": "nomic-embed-text" }
  }
}

Standalone CLI

npm install -g clawmem
clawmem init
clawmem add "I prefer TypeScript over Python"
clawmem search "programming preferences"
clawmem profile

---

Packages

Package	Description	npm
@clawmem/core	Core memory engine (extraction, vector+graph, dedup, profiles)	npm i @clawmem/core
@clawmem/openclaw	OpenClaw memory-slot plugin	openclaw plugins install @clawmem/openclaw
clawmem	Standalone CLI	npm i -g clawmem

---

Storage Backends

Layer	Default	Alternatives
Vector store	sqlite-vec	LanceDB, Qdrant
Graph store	Kùzu (embedded)	Neo4j
History	SQLite	—
LLM	OpenAI-compatible	Any
Embedder	OpenAI-compatible	Any

All defaults are embedded — no servers, no Docker, no accounts.

---

Safety by Default

• No outbound calls unless you explicitly configure a remote endpoint
• Full audit log: every memory mutation records previous value + diff
• Reversible: clawmem history <id> shows version chain, clawmem revert <id> <version> restores
• Group-safe: auto-skips capture in OpenClaw group chats
• Hard caps: configurable max memories, tokens, retention days

See THREAT-MODEL.md for the full security model.

---

Architecture

packages/core          @clawmem/core
  ├─ Memory engine     add / search / get / getAll / delete / update
  ├─ Extraction        LLM → structured facts (13 categories)
  ├─ Deduplication     hash + semantic + LLM merge
  ├─ Graph memory      Kùzu: entities, relationships, UPDATE/EXTEND/DERIVE
  ├─ User profiles     static (identity/prefs) + dynamic (projects/goals)
  ├─ Sleep mode        nightly consolidation + digest
  └─ Pluggable backends

packages/openclaw-plugin   @clawmem/openclaw
  ├─ Memory-slot plugin    plugins.slots.memory
  ├─ 7 tools               search, store, store_raw, list, get, forget, profile
  ├─ Auto-recall           injects memories before each agent turn
  ├─ Auto-capture          extracts facts after each agent turn
  ├─ Markdown sync         two-lane: workspace .md ↔ vector+graph store
  └─ Identity mapping      multi-channel user resolution

packages/cli               clawmem
  └─ Standalone CLI        init, add, search, forget, profile, graph, sleep, doctor

---

Development

# Prerequisites: Node 20+, pnpm 9+
git clone https://github.com/DeepExtrema/clawmem
cd clawmem
pnpm install
# If any package.json/workspace dependency metadata changes:
# rerun pnpm install and commit pnpm-lock.yaml in the same PR
pnpm build
pnpm test          # 130+ tests across core, CLI, and plugin

Optional: Enable ANN vector search

# Install sqlite-vec for fast approximate nearest neighbor search
# Without it, ClawMem falls back to O(n) cosine similarity (fine for <10k memories)
pnpm add sqlite-vec --filter @clawmem/core

Error Types

ClawMem exports typed errors for structured error handling:

import { ClawMemError, LLMError, EmbedderError, StorageError } from "@clawmem/core";

try {
  await memory.add(messages, { userId: "user1" });
} catch (err) {
  if (err instanceof LLMError) {
    // LLM timeout, HTTP error, or empty response
  } else if (err instanceof EmbedderError) {
    // Embedding dimension mismatch, timeout, or HTTP error
  } else if (err instanceof StorageError) {
    // Database error
  }
}

See CONTRIBUTING.md for contribution guidelines and docs/ for full API reference.

---

License

Apache 2.0 © 2026 DeepExtrema