TotalReclaw
End-to-end encrypted memory + knowledge graph for AI agents — portable, yours forever
Website ·
npm (plugin) ·
npm (MCP) ·
PyPI ·
Getting Started ·
Architecture
Status: private beta — free tier open, no credit card, no waitlist.
Your AI remembers everything — without remembering it for Big Tech. And starting with v1, every memory is structured: 6 speech-act types, provenance, scope, and volatility on every entry.
- Private — XChaCha20-Poly1305 encryption happens on-device. The server never sees plaintext, even if fully compromised.
- Structured — every memory typed by speech act with source, scope, and volatility. Learn more →
- Portable — one 12-word recovery phrase works across every supported agent. Switch platforms without losing a single memory. Compare integrations →
- Yours forever — memories are anchored to Gnosis Chain and indexed by The Graph. No vendor lock-in, no data hostage.
Most AI memory products store your data unencrypted server-side, or leave encryption up to your deployment. TotalReclaw can't read your data by design — that's the difference between a privacy promise and a structural guarantee.
|
TotalReclaw |
Typical hosted memory |
| Server sees your memories in plaintext? |
Never — E2E encrypted on-device |
Yes |
| One-click plain-text export? |
Yes |
Varies |
| Works across agents? |
Yes — OpenClaw, Claude Desktop, Hermes, ZeroClaw, any MCP host |
Usually SDK-locked |
| Storage you control? |
Yes — public chain or your own PostgreSQL |
Vendor database |
New in v1 (April 2026): memory taxonomy (6 types), source-weighted reranking, 4 new MCP tools (totalreclaw_pin / totalreclaw_unpin / totalreclaw_retype / totalreclaw_set_scope), protobuf v4 wire format. v1 is the default on every client. Existing vaults decrypt transparently. See the v1 migration guide.
Quick Start
One command per client. v1 is the default — no env toggles, no feature flags.
Prefer the chat-driven install? Paste one host-agnostic line into any agent (OpenClaw, Hermes, Claude Desktop / Cursor / Windsurf): Install TotalReclaw. See https://totalreclaw.xyz/install — a short alias for the universal dispatcher (install.md).
| Client |
Install |
| OpenClaw |
openclaw plugins install @totalreclaw/totalreclaw |
| Claude Desktop (and other MCP chat apps) |
Add npx -y @totalreclaw/mcp-server to the app's MCP config, restart, then ask "Set up TotalReclaw" — see the Claude Desktop guide |
| Claude Code / Cursor / Windsurf |
claude mcp add -s user totalreclaw -- npx -y @totalreclaw/mcp-server (or the JSON config), then ask "Set up TotalReclaw" — see the coding-tools guide |
| Codex (CLI, IDE extension, ChatGPT desktop app) |
codex plugin marketplace add graphprotocol/totalreclaw then codex plugin add totalreclaw@totalreclaw, restart, then ask "Set up TotalReclaw" — see the Codex guide |
| NanoClaw |
@totalreclaw/skill-nanoclaw (bundled via NanoClaw) + set TOTALRECLAW_RECOVERY_PHRASE in deployment env — see the NanoClaw guide |
| Python / Hermes |
pip install totalreclaw |
| Rust / ZeroClaw |
cargo add totalreclaw-memory |
| IronClaw |
via the generic MCP server (first-class integration paused) — see IronClaw setup |
Then set one env var on your host:
export TOTALRECLAW_RECOVERY_PHRASE="your twelve word recovery phrase here"
If you don't have a phrase yet, the setup wizard generates one. See the client setup guide for per-platform details.
How It Works
- Your agent extracts facts from conversations — typed by speech act (claim, preference, directive, commitment, episode, summary) and tagged with source, scope, volatility.
- Everything is encrypted on-device — keys derived from your recovery phrase, never sent anywhere.
- Encrypted data is stored on-chain — the server only sees ciphertext and hashed search tokens.
- Search works over encrypted data — blind indices and LSH let the server find relevant memories without reading them.
- Results are decrypted and ranked locally — BM25 + cosine + RRF with source-weighted reranking (v1 Tier 1).

Full technical deep dives: Architecture, Memory Taxonomy v1, Tiered Retrieval.
What's new in v1
v1 shipped across every client in April 2026. Highlights:
- 6-type speech-act taxonomy —
claim / preference / directive / commitment / episode / summary. Replaces the ambiguous 8-type v0 list.
- 3 orthogonal axes —
source (user / user-inferred / assistant / external / derived), scope (work / personal / health / family / creative / finance / misc / unspecified), volatility (stable / updatable / ephemeral).
- Source-weighted reranking — user-authored claims consistently rank above assistant-regurgitated noise. Structurally fixes the 97.8% junk problem documented in other memory systems.
- 4 new MCP tools —
totalreclaw_pin, totalreclaw_unpin, totalreclaw_retype, totalreclaw_set_scope. Invoked by natural language ("pin that", "file that under work").
- Protobuf v4 wire format — outer wrapper bump; subgraph schema unchanged.
- Env var cleanup — several experimental env vars removed; a small set of user-facing env vars remain. See the env vars reference.
Full details: v1 migration guide and memory types guide.
Import from Other AI Tools
Already using another AI assistant with memory? Tell your agent:
- "Import my ChatGPT memories into TotalReclaw"
- "Import my Claude memories into TotalReclaw"
- "Import my Gemini history into TotalReclaw"
Supported sources: ChatGPT (memories or conversations.json), Claude (memory export), Gemini (Google Takeout), Mem0 (API or file), MCP Memory Server (JSONL). All imports are encrypted client-side before storage. See the import guide.
On the managed service, the free tier includes one lifetime import; re-importing from additional sources is a Pro feature. See pricing.
Your vault in the browser
View and curate your memories at app.totalreclaw.xyz — no install required. Your recovery phrase is entered locally and never leaves your device; decryption happens entirely in the browser, so the server only ever serves ciphertext.
Packages
| Package |
Description |
Install |
| @totalreclaw/totalreclaw |
OpenClaw plugin — automatic encrypted memory |
openclaw plugins install @totalreclaw/totalreclaw |
| @totalreclaw/mcp-server |
MCP server for Claude Desktop and other MCP clients |
npx @totalreclaw/mcp-server setup |
| @totalreclaw/client |
Client library (encryption, indexing, search, re-ranking) |
npm install @totalreclaw/client |
| totalreclaw |
Python client + Hermes Agent plugin |
pip install totalreclaw |
Pricing
Free: 250 memories/month, permanently stored on Gnosis mainnet, end-to-end encrypted — no credit card required. Pro: 3,000 memories/month, plus LLM-guided dedup and custom extraction intervals. See pricing →
Self-Hosting
Run the open-source server with PostgreSQL — no dependency on totalreclaw.xyz:
cd server && cp .env.example .env && docker-compose up -d
Set TOTALRECLAW_SELF_HOSTED=true and TOTALRECLAW_SERVER_URL=http://localhost:8080 on the client. Both modes encrypt identically — the difference is where the encrypted blobs are stored.
Repository Structure
totalreclaw/
├── client/ TypeScript client library (encryption, indexing, search, re-ranking)
├── rust/ Rust core (XChaCha20-Poly1305, LSH, reranker) — compiles to WASM + PyO3
├── python/ Python client library + Hermes Agent plugin
├── skill/ OpenClaw plugin (automatic memory via lifecycle hooks)
├── skill-nanoclaw/ NanoClaw skill package + MCP bridge
├── mcp/ MCP server for Claude Desktop and other MCP clients
├── app/ Vault SPA (Vite + React) — managed-mode read UI
├── server/ Self-hosted server (FastAPI + PostgreSQL)
├── contracts/ Smart contracts (Gnosis Chain)
├── subgraph/ The Graph subgraph (AssemblyScript indexer)
├── docs/ Architecture, specs, and guides
└── tests/ Integration, parity, and E2E tests
Documentation
Start here
Per-client setup
Specs + architecture
Other
License
- Self-hosted server (
server/) — AGPL-3.0
- All other code — MIT