AI agents can now buy memory.
A paid memory API where agents store, recall, and delete memory for $0.001 per call using USDC on Base. No signup. No API key. No account.
Watch the transaction happen live ↓
$ click Run demo to watch a server-side wallet pay $0.001 USDC for a real /api/memory/store call. No signup, no API key — only a wallet signing a typed-data hash.
What just happened
- 1.The server probed
POST /api/memory/storeand got back402 Payment Requiredwith payment details (amount, payTo, asset, USDC EIP-712 domain). - 2.A throwaway wallet signed an EIP-3009
TransferWithAuthorizationgranting permission to move 1000 atomic USDC ($0.001) to the recipient. - 3.The signed payload was base64-encoded into an
X-PAYMENTheader and the request was replayed. - 4.Edge middleware decoded the payment, forwarded it to the Coinbase CDP facilitator, which verified the signature and broadcast the USDC transfer on Base.
- 5.
memory_storeran, persisted the row, and the response came back. ~10s later the transfer is indexed on Basescan.
Why this matters
AI agents need to consume APIs autonomously. Today that means provisioning API keys, signing up for accounts, managing billing per-vendor — none of which an LLM can do unattended. With x402, an agent points at a URL, gets a 402, signs a transfer with its own wallet, and access is granted. No accounts, no per-vendor integration, no chargebacks. The same wallet works across every x402-compatible service, indexed at agentic.market.
Want to use the Agent Memory API yourself? Run the MCP server in your client of choice — see /memory for endpoints, MCP install instructions, and the full protocol spec.
The real question
If agents can spend money, companies need controls.
Memory is just one paid API. Multiply this demo by every API your agent stack will consume — search, inference, scrapes, vector DBs. A single rogue loop drains your wallet in minutes. The next problem isn't whether agents can pay. It's budgets, per-tool caps, kill switches, and spend visibility per agent.
See AgentGuard — runtime spend controls for AI agents →