an-am-NEE-sis — from the Greek, meaning "unforgetting."
Sovereign AI memory. Yours forever.
Read the technical specification →
Anamnesis is an open protocol for persistent, encrypted, user-owned AI context. Your conversation history, knowledge, and relationship with AI lives on infrastructure you control — not on a company's server.
When you talk to an AI, it forgets you the moment the session ends. You start over every time. Anamnesis is the fix: your context persists across every session, every model, every platform. Encrypted with a key derived from a token you own. Readable only by you.
This is a basic human right — the right to your own memory.
If OpenAI shuts down tomorrow, your context is on Hedera. Plug it into the next provider. Nothing lost. Your AI knows you the same way it did yesterday — because the memory was never theirs to begin with.
The same encrypted vault works with any provider. Use GPT today, Claude tomorrow, a model that doesn't exist yet in five years. Your AI relationship is not tied to any company. It travels with you.
Your accumulated context — everything your AI has learned about you, your domain, your work — is transferable. Sell your expertise. Gift your knowledge to a successor. Pass your AI relationship to your heirs. This is digital inheritance. It has never existed before.
Anamnesis includes agentic tools: your AI can retrieve your passwords, create calendar events, send encrypted messages, manage your files — all within cryptographic session boundaries. It acts on your behalf without any platform seeing the data it touches.
For journalists, whistleblowers, and researchers working in sensitive territory: the protocol includes a dead man's switch. A smart contract automatically transfers your encrypted vault to a pre-designated journalist or organization if you stop checking in. If you are arrested, disappeared, or killed — the documentation reaches the people you chose. No intermediary. No one can stop it from firing.
Every context update is logged to Hedera Consensus Service with a timestamp and content hash. Your AI's memory is not just private — it's verifiable. Cryptographic proof that no one has touched it.
When an employee leaves, their accumulated AI context — their domain knowledge, their reasoning patterns, their work history — can be retained and transferred with their consent. Not lost, not stolen. Transferred.
Every AI platform today can read your context, train on it, sell it, or hand it to a government. With Anamnesis, there is no plaintext on any server. The provider receives your context for one session and nothing more. Your AI relationship is yours — not a service you rent from a company that owns everything you tell it.
Every AI platform today owns your context:
- Your history lives on OpenAI's servers, Google's servers, Anthropic's servers
- They can read it, train on it, sell it, or delete it
- Switch providers and lose everything
- Platform shuts down — your AI relationship is gone
- No cryptographic proof your data hasn't been tampered with
This is not how it should work.
Anamnesis solves this with three components:
Your encryption key is never stored anywhere. It is derived fresh each session using HKDF-SHA256 from:
key = HKDF(token_id + wallet_signature)
The token is the proof of ownership. The wallet signature is the proof of presence. No one else can produce this key — not the server, not us, not anyone.
This mechanism is chain-agnostic. Any chain that supports token ownership can serve as the ownership layer. The reference implementation uses Hedera, but the protocol is not Hedera-specific.
Encrypted context lives on Hedera via HCS-1 — an immutable, decentralized ledger. It survives:
- Platform shutdowns
- Terms of service changes
- Company acquisitions
- Government subpoenas (no plaintext to hand over)
Every context update is logged to Hedera Consensus Service (HCS) with a timestamp and content hash — a permanent, tamper-proof audit trail.
The same encrypted vault works with any AI provider:
- OpenAI, Anthropic, Google, Mistral, Groq, Ollama
- Switch models without losing context
- Bring your own API keys
- Your context is injected identically to every provider
┌─────────────────────────────────────────────────────┐
│ Your Device │
│ │
│ wallet_signature + token_id │
│ │ │
│ ▼ │
│ HKDF-SHA256 ──► AES-256-GCM key (never stored) │
│ │ │
│ ▼ │
│ Decrypt context ◄──────────────────────────────┐ │
│ │ │ │
│ ▼ │ │
│ [AI Provider of your choice] │ │
│ │ │ │
│ ▼ │ │
│ Re-encrypt ────────────────────────────────────┘ │
└─────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ Hedera Hashgraph │
│ │
│ HTS — Token ownership proof │
│ HCS-1 — Encrypted context storage │
│ HCS — Immutable audit log │
│ EVM — Access control smart contracts │
└─────────────────────────────────────────────────────┘
- Wallet signs a server-issued challenge
- HKDF-SHA256 derives AES-256-GCM key from
token_id + signature - Encrypted context downloaded from HCS-1 and decrypted locally
- Plaintext never leaves your device — only injected into AI provider call
- After session: re-encrypted, written back to HCS-1, update logged to HCS
The reference implementation uses Hedera (HTS + HFS + HCS). Any stack that provides token ownership, decentralized storage, and an immutable audit log can implement this protocol.
| Layer | Reference | Alternatives |
|---|---|---|
| Token ownership | Hedera HTS | Ethereum ERC-721, Polygon, Solana, Base |
| Encrypted storage | Hedera HCS-1 (recommended) | Arweave (permanent, pay-once), IPFS + Filecoin, Hedera HFS |
| Audit log | Hedera HCS | Any chain event log, Arweave |
Arweave is worth particular consideration for the storage layer — pay once, stored permanently, genuinely decentralized with no governance council. A community implementation using Ethereum + Arweave would be a meaningful contribution to the ecosystem.
If you build an implementation on a different stack, open a PR to add it to this list.
This repository is the reference implementation:
anamnesis/
├── api/ # FastAPI backend
│ ├── main.py
│ ├── routes/ # auth, context, companion, suite endpoints
│ └── services/ # Hedera, encryption, RAG services
├── src/ # Core protocol
│ ├── crypto.py # HKDF + AES-256-GCM
│ ├── context_token.py # HTS token minting
│ ├── vault.py # HFS storage
│ └── memory_packages.py # Modular context / RAG
├── frontend/ # React web interface
├── frontend-src/ # Tauri desktop app (reference implementation)
├── mobile/ # Expo React Native app
├── contracts/ # Solidity smart contracts (Hedera EVM)
├── scripts/ # Admin and setup tools
└── tests/ # Test suite
The reference implementation includes a full encrypted personal suite, all keyed to the same token:
| Service | What it stores |
|---|---|
| AI Companion | Persistent conversation context |
| Pass | Passwords, secure notes, TOTP codes |
| Drive | Encrypted file storage |
| End-to-end encrypted messaging between token holders | |
| Calendar | Encrypted calendar events |
| Knowledge | Personal knowledge base, RAG-indexed for AI context |
New here? → QUICKSTART.md — 15 minutes from zero to first working implementation.
- Python 3.10+
- Hedera testnet account — free at portal.hedera.com
git clone https://github.com/anamnesis-protocol/anamnesis.git
cd anamnesis
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
# Add your Hedera testnet credentials to .envuvicorn api.main:app --reload --port 8000
# Open http://localhost:8000/demopytest tests/ -v
# Crypto tests only (no network required)
pytest tests/test_crypto.py -v| Layer | Technology |
|---|---|
| Backend | FastAPI (Python 3.11+) |
| AI Providers | OpenAI, Anthropic, Google, Mistral, Groq, Ollama |
| Desktop | Tauri (Rust + React + TypeScript) |
| Mobile | Expo (React Native) |
| Blockchain | Hedera (HTS, HFS, HCS, EVM) |
| Encryption | AES-256-GCM, HKDF-SHA256, WebAuthn PRF |
| Database | Supabase (optional BYOK storage) |
| Build | Vite, TypeScript 5 |
Arty Fitchels is the hosted consumer implementation of this protocol — a managed service for users who want sovereign AI context without running their own infrastructure.
The protocol is open. The hosted service is separate.
Anamnesis is a public good. Contributions are welcome.
See CONTRIBUTING.md for guidelines.
Areas where help is most needed:
- Additional AI provider integrations
- Mobile apps (iOS/Android)
- Browser extension
- Alternative storage backends (IPFS, Arweave)
- Client libraries (JS, Rust, Go)
- Documentation and translations
From the Greek — unforgetting. Plato's concept that learning is the recovery of knowledge the soul already possessed.
Your AI context was always yours. We just gave it back.
The full protocol specification — cryptographic primitives, session protocol, storage layer, ownership layer, audit trail, and prior art statement — is in WHITEPAPER.md.
Apache License 2.0 — see LICENSE.
This means you can use, modify, and distribute this freely — commercially or otherwise — with no royalties and no restrictions beyond attribution. No one, including the original authors, can ever patent-trap you on this protocol.