MnemoPay gives autonomous agents persistent memory, real payment rails, signed receipts, portable reputation, and human approval gates before they act in the real world.
MnemoPay gives AI agents the missing trust layer: identity, memory, payment rails, spend policy, signed receipts, and audit bundles on one regulator-shaped chain.
Every agent gets a keypair. Every action is signed. Every receipt is independently verifiable. Canary honeypots flag impersonation in real time.
Every fact the agent stores is Merkle-chained. Tampering leaves a cryptographic trace. Decay policies, recall ranking, PII redaction, all built in.
Stripe, Paystack, Lightning — plus Stripe MPP, x402 (USDC on Base), and Google AP2, all stable in v1.6.0. Same agent, same score, same audit chain across every rail.
Capability layer the agent uses to do things — runtime, browser via CDP, computer-control, vision, voice, scrape, knowledge. Charter-bound, FiscalGate-aware, audit-logged.
Reserve the budget on charter start. Every priced action routes through hold → settle. Over-budget halts. Errors release. Agents never spend what they didn't declare.
Merkle-rooted, signed. Every tool call, payment, memory write, refusal. Mission · events · CSV · chain · manifest with checksums. Export-ready for any regulator.
“Mission → authority check → tools → proof.”
The same six steps run for every priced action your agent takes — deterministic, replayable, and signed end-to-end.
A goal arrives — from a user, a queue, an upstream agent. The runtime accepts it and instantiates a session.
YAML / JSON declares the rules: goal, budget cap, allowed tools, allowed counterparties, approval thresholds, retention.
The budget is reserved on charter start. Every priced action hits hold → settle. Over-budget halts. Errors release.
Cheapest viable rail is selected automatically — Stripe, Paystack, Lightning, MPP, x402, AP2 — against the agent's score.
Ed25519-signed receipt is produced and chained into the agent's Merkle log. The counterparty can verify independently.
Article 12 bundle is appendable on demand: mission · events · CSV · chain · manifest with checksums. Regulator-shaped.
Every run is replayable. Same charter, same inputs, same chain root.
No tool call escapes Charter + FiscalGate. The runtime refuses what the charter doesn't allow.
One identity, one score, one chain — across every rail the agent ever touches.
Humans had identity, payroll, and receipts.
Agents need the same trust layer.
MnemoPay gives money-moving agents
memory, rails, reputation, and audit trails.
Agent trust is not a single category — it is the same money, memory, and audit problem hitting four different desks at the same time.
Running five agent SKUs across Claude, GPT, and open-source. Auditors keep asking "how do you know what the agent did?" — and you don't have a clean answer.
Agents now sign POs, refund customers, and pay invoices. The card-on-file model from 2024 has no per-agent attribution, no per-agent cap, no per-agent receipt.
Onboarding third-party agents whose reputation you cannot verify. Reference letters in PDF don't scale; you need a portable credit score that travels across customers.
August 2026 is when Article 12 logging requirements bite for general-purpose AI systems. You can either retrofit logging across three vendor stacks, or deploy AI tools that ship the personnel file by default.
The SDK is Apache-licensed and free forever. You pay when the personnel file leaves your machine — managed rails, hosted scores, regulator-shaped audit bundles, or enterprise SLAs. Lightning settles at network fee — we take nothing on top.
For Chief Compliance Officers staring down the August 2026 Article 12 logging deadline. Every agent action in your stack — across MnemoPay, your existing tools, or both — bundled, Merkle-rooted, Ed25519-signed, and export-ready in the shape the EU AI Office actually accepts. We run the audit chain so your CCO doesn't have to.
W-2, reference check, employment contract, corp memory, audit-ready personnel file — in two lines of code. One SDK. Apache 2.0.
