SealedMind

What is SealedMind?

SealedMind is a privacy-first AI memory primitive built on the 0G stack. It gives every user a sovereign, encrypted, AI-searchable memory vault — called a Mind — that nobody, not even the application operator, can read without explicit, on-chain consent. Think of it as a personal second brain that lives on a decentralized network, processed inside a hardware-level trusted execution environment, owned entirely by the user's wallet, and shareable with other agents through a revocable on-chain capability. Live on 0G mainnet (chainId 16661) and Galileo testnet (16602) today.

The Problem

Every useful AI agent — your coding copilot, your therapist bot, your trading assistant — gets smarter with memory. But today that memory is broken in three ways. One, it doesn't persist across agents — you explain yourself to ChatGPT, then to Claude, then to your company's internal AI. Two, when AI does remember, you don't own it — your history belongs to OpenAI's servers, you can't take it with you, can't sell it, can't will it to your kid. Three, there is zero privacy proof — "we don't read your data" is a policy, not math, and there's no cryptographic receipt that the LLM ever processed your data inside a sealed environment. The whole AI-agent category is stuck because agents have no property semantics.

The Solution

SealedMind composes four pieces of the 0G stack into a single privacy primitive. Every memory is encrypted client-side with AES-256-GCM under a key derived from the user's wallet signature via HKDF — the server is mathematically blind to plaintext. The encrypted ciphertext is uploaded to 0G Storage and content-addressed by rootHash. On 0G Chain, an ERC-7857 iNFT called a Mind NFT represents the user's full memory vault — transferable, composable, owned by the wallet. When a user recalls memory, the query is embedded locally (all-MiniLM-L6-v2, 384-dim), HNSW vector search finds the top-K nearest memories, and the matching content is sent into 0G Sealed Inference — Qwen 2.5 7B running inside Intel TDX with a confidential NVIDIA H100 GPU pass-through. The plaintext never exists outside the enclave, and every reply comes back with a hardware-signed TEE attestation. Crucially, every remember, recall, and chat operation now fires a background transaction to our on-chain MemoryAccessLog — and the chainscan link is returned in the verify response. The "Verify Proof" button on every reply isn't a checkmark, it's a link to the on-chain truth.

Smart Contracts

SealedMindNFT is an ERC-7857 intelligent NFT where one token equals one user's entire memory vault. It stores the 0G Storage CIDs, memory count, shard names, and authorized user list per Mind. Transferable — you can sell, lease, or inherit your Mind.

CapabilityRegistry lets Mind owners grant fine-grained access to specific memory shards — for example, granting a doctor's AI read-only access to the health shard for 30 days while keeping finance and relationships private. Capabilities are on-chain, time-expiring, scope-limited, and revocable in a single transaction. Instant 403 for the grantee on the next read.

MemoryAccessLog is an immutable, on-chain audit trail. Every memory operation — remember, recall, chat — emits an accessRecorded event with the mindId, operation, attestation hash, and storage CID. Wired end-to-end into the backend so the chainscan transaction hash returns in the verify response, making every access independently auditable in one click.

Verifier is the TEE attestation verifier wired into SealedMindNFT for hardware-attestation validation.

Backend

The backend runs on Node.js plus Express plus TypeScript, hosted on Railway. It handles Sign-In with Ethereum authentication, issues long-lived API keys for end-users and operator integrations, performs two-pass fact extraction (regex fast-path then TEE-LLM fallback for ambiguous cases), encrypts memories with AES-256-GCM under HKDF-derived per-user keys, generates 384-dimensional embeddings using HuggingFace all-MiniLM-L6-v2, stores encrypted blobs on 0G Storage via the official 0G TypeScript SDK, maintains an HNSW vector index for fast cosine-similarity search, routes recall queries through 0G Sealed Inference via the 0G Serving Broker, fires the background MemoryAccessLog transaction for every operation, and exposes a standalone TEE-attested chat endpoint for any agent that just needs sealed inference.

Frontend

The React 19 dApp uses Vite, Tailwind v4, RainbowKit, wagmi, and viem for wallet connectivity. Users can mint their Mind iNFT, store memories into named shards (health, finance, work, relationships, general), recall using natural-language queries, and grant or revoke shard-level capabilities to other wallet addresses — all on-chain. Every recall response renders an AttestationCard showing the TEE attestation chip plus a "Verify Proof" button — when clicked, the backend re-checks the attestation chain and returns the on-chain MemoryAccessLog chainscan link. The proof is a click, not a claim. The site also includes a live architecture page (full SVG diagram of the 6 trust boundaries plus a 6-card threat model), a developer self-serve onboarding page (wallet, SIWE, API key in 30 seconds), a demo page with a live two-agent capability demo plus copy-paste scripted prompts and a self-serve test guide, an access landing page for shared capabilities, and a deck presentation route for the pitch.

CLI and Agents

The sealedmind CLI ships with login, remember, recall, grant, and revoke commands — scriptable JSON output, persistent API key after one-time SIWE login. A Life OS agent built on OpenClaw and Claude Haiku 4.5 lets users interact with their Mind through natural conversation in the terminal. We also built a live two-agent demo — Aria (Alice's personal assistant on Claude) talks to Dr. Chen's clinical AI (Claude orchestration plus Qwen 2.5 7B in Intel TDX for the privileged synthesis). The demo runs the full capability grant, recall, revoke lifecycle live in a browser, with every on-chain transaction surfaced in a live event feed.

Three SDKs Published

@sealedmind/sdk on npm — TypeScript, MIT-licensed, full typings, three method calls (remember, recall, grantCapability), works in any JavaScript runtime.

sealedmind on PyPI — async Python SDK via httpx, same surface area, for LangGraph, smolagents, AutoGen, CrewAI, OpenClaw, or any custom Python agent.

evermemos-sealedmind on PyPI — a sanctioned drop-in addon for 0gfoundation/0g-memory, 0G's own official memory project. One pip install, one env var, and any existing 0G Memory deployment gets encrypted-under-wallet-key memory plus on-chain audit. This is the move that makes us a contributor to the 0G stack itself, not just a consumer of it.

Partnerships

Daimon — a consumer dApp for tradeable AI trading agents. Every Daimon's brain is a SealedMind ERC-7857 iNFT. Memory under the owner's wallet key, capability rental for renting access to the brain, on-chain audit on every read. Marketplace contract live on both Galileo testnet AND 0G mainnet at 0xb9D42824955b492BE4cBf13988C3d0Ad9985F807. Without SealedMind, Daimon doesn't exist.

VeilSolver — the MEV-resistant intent solver on 0G. They had a bespoke encrypted-storage layer for trader strategies and audit blobs. They ripped that out and replaced it with @sealedmind/sdk calls — their strategy registry, audit trail, and compliance log all run on SealedMind now. Joint integration guide published in our repo.

We talked with over ten teams in the 0G ecosystem about integrating. Two went all the way and shipped on top of us. We've been present at every 0G builder showcase and every builder meet since day one of the program.

0G Technologies Used

0G Storage for all encrypted memory blob persistence (content-addressed by rootHash). 0G Sealed Inference for both fact extraction during remember and synthesis during recall — Qwen 2.5 7B in Intel TDX plus H100, attested per call. 0G Chain (mainnet 16661 plus Galileo testnet 16602) for all smart contract interactions — Mind iNFT minting, capability grants and revocations, and the immutable MemoryAccessLog audit trail. 0G Memory — we extended it directly via our evermemos-sealedmind PyPI addon, hooking into the official memsys.addons entry point.