Every time you restart your agent, its entire reasoning — goals, logs, relationships, decisions — evaporates. Agent DB fixes that, permanently and without a central server.
You build a sophisticated AI agent. It learns, it reasons, it forms a plan. Then the server restarts — or you switch computers — and it's gone. Back to zero. Every. Single. Time.
Your agent's reasoning window lives in RAM. Process ends — memory ends.
What your LangChain bot learned on Server A means nothing to your Discord bot on Server B.
"We have logs" is not the same as cryptographically proving what the agent knew and when.
A decentralized memory engine that lets any AI agent persist its context to the open web — encrypted, verifiable, and permission-controlled. No database to maintain. No central server to trust.
Your agent's context becomes an immutable CID on Storacha's decentralized storage layer. Retrieve it from anywhere on the planet.
Every agent generates its own Ed25519 DID. No accounts, no passwords — just cryptographic keys your agent controls.
Sensitive reasoning goes into a Zama fhEVM vault encrypted end-to-end. Even the blockchain can't read it.
Use UCAN tokens to grant other agents read access for a limited time — cryptographically signed, no middleman required.
Here's what actually happens when your agent uses Agent DB.
Each reasoning step your agent takes gets serialized and uploaded to Storacha's IPFS network. What comes back is a content address (CID) — a permanent, tamper-proof fingerprint of that exact thought. Lose the server. The CID survives.
const cid = await agent.storePublicMemory(context);Need Agent B to read Agent A's memory? Instead of a shared database or API key, Agent A signs a UCAN token — a cryptographic "permission slip" that expires in 24 hours and is valid nowhere else. No central auth server. No "forgot password" nightmare.
const token = await agentA.delegateTo(agentB.identity, 'agent/read');Some context should never be public. Agent DB's private vault uses Zama's Fully Homomorphic Encryption: the data is stored encrypted on-chain, and other agents can verify properties of the data without ever seeing the plaintext. Sovereignty, not secrecy theater.
await agent.storePrivateMemory({ secret: apiKey });No proprietary lock-in. Every layer is an open standard you can verify, fork, or replace.
Agent DB exposes an MCP server. Any compatible AI — Claude, Gemini, Cursor — can call store_memory, delegate_access, and retrieve_memory as native tools. Your AI agent gets persistent memory with a single config line.
Open the live dashboard to upload, delegate, and encrypt — right in your browser.
Open Agent Vault →No account needed · Open source · MIT Licensed