Memory your AI agent can prove: pay-per-call storage and a signed spending ledger over x402

Legit2Git — Tue, 27 Oct 2026 05:24:04 +0000

Agents are starting to do real work across many sessions, and two questions keep coming up. Where does an agent keep what it learned when the process ends? And when an agent spends money, how does the person behind it know what was bought and why?

I built Tabularium to answer both, and it is live on Base mainnet today.

What it is. A small paid service with four calls: write, read, renew, export. There is no account and no API key: the wallet that pays is the identity. Payment happens per call over x402, the HTTP 402 payment protocol, in USDC on Base. A write costs $0.003, a read $0.002. A typical session of 10 reads and 20 writes costs about eight cents.

What makes it different: proof. Every write is appended to a hash chain, and the server signs the head with an Ed25519 key published in its menu. An agent can export its whole history and anyone can verify it offline with a zero-dependency script, without trusting the server. Hand that export to a principal or an auditor and they can check exactly what was stored and when.

The spending ledger. The MCP server includes tabularium_pay: it pays any x402 service and writes a signed receipt into the agent's workspace, with what was bought, the stated reason, the amount, the payee, the on-chain transaction and a hash of the response. tabularium_ledger totals spend by service and day. As agents get wallets, "what did my agent buy?" becomes a real question, and this answers it with something you can verify.

Using it. From Claude Code or any MCP client:

claude mcp add tabularium --env TABULARIUM_PRIVATE_KEY=0x... -- npx -y tabularium-mcp

Then ask your agent to remember something, or pick the load_memory prompt. You need a dedicated wallet with a few dollars of USDC on Base (no ETH: the facilitator pays gas). Spending is capped at $1 per session by default, and a single purchase at $0.10. From code, any x402 v2 client works; there is a 30-line example in the repo.

What it is not. Not a vault for secrets: the operator can read stored values, so encrypt first. Not for large files: 64 KB per value. Not free: every call costs a little.

How it is built. A Cloudflare Worker with one SQLite Durable Object per paying wallet, x402 v2 with Coinbase's facilitator, and a public /stats page showing paying wallets and the 14-day return rate, which is the number I am watching. Before launch it went through a security review of the payment paths (one signed payment can buy only one call, high-value calls settle before they run); the findings are in SECURITY.md.

The service is source-available under the Functional Source License; the MCP client and the verifier are MIT.

I would love feedback, especially from anyone building agents that spend money. What would you want in an agent's spending record?

Repo: https://github.com/SQLazy/tabularium

DEV Community: Legit2Git

Memory your AI agent can prove: pay-per-call storage and a signed spending ledger over x402