MnemoPay

The governance layer for AI agents that handle money. Charter-driven mission scope, FiscalGate budget enforcement, EU AI Act Article 12 audit bundles, Agent Reputation Scoring (300-850), and a tamper-evident MerkleAudit chain — across every payment rail an agent will ever touch.

MnemoPay sits above the rail (Stripe, Paystack, Lightning, Stripe MPP, x402, Google AP2) and below the agent runtime (LangChain, CrewAI, Claude Agent SDK, your own loop). The rail moves money. The runtime decides. MnemoPay declares the rules, enforces the budget, and produces the evidence.

npm install @mnemopay/sdk

New here? Start at docs/QUICKSTART.md — 60 seconds, three steps, working code.

Docs: Quickstart · Architecture · Permissions · Action ledger · Integrations (OpenAI/Anthropic/Gemini/Cohere/Mistral/LangGraph) · Recall · FiscalGate · Audit bundles (EU AI Act Art. 12) · Subpath import rule · Claude Agent SDK guide · Bundlers: Vite · Webpack · Bun

Community: LICENSE (Apache 2.0) · CHANGELOG · CONTRIBUTING · CODE_OF_CONDUCT · SECURITY · Discussions · Good first issues

Receipts: Trust hub (entity, KYB, Apple Team ID, Article 12 audit chain — verify in <5 min) · Benchmarks (1M ops, 100% adversarial detection, $0 ledger drift) · Python SDK on PyPI (full TS-rail parity since 1.1.0)

import MnemoPay, {
  Charter, FiscalGate, MerkleAudit,        // governance primitives
  AgentReputationScoring, BehavioralEngine, // trust + reputation
  StripeRail, X402Rail, GoogleAP2Rail,      // rails
} from "@mnemopay/sdk";

const agent = MnemoPay.quick("my-agent");

await agent.remember("User prefers monthly billing");
const tx = await agent.charge(25, "Monthly API access");   // FiscalGate hold
await agent.settle(tx.id);                                  // FiscalGate capture

// Agent Reputation Score — portable, 300-850 range. NOT FICO-brand, NOT a consumer
// credit report, NOT governed by FCRA. Scores agents (software), not humans.
const scorer = new AgentReputationScoring();
const result = scorer.compute({ transactions: [tx], createdAt: new Date(), /* ... */ });
// → { score: 672, rating: "good", feeRate: 0.015, trustLevel: "standard" }

14 modules. Hash-chained ledger. Charter / FiscalGate / Article 12 audit bundles. 6 payment rails. 200K-operation stress tested. Apache 2.0.

What MnemoPay is NOT: not a bank, not a money transmitter, not a Stripe replacement, not an agent framework, not a compliance platform. It's the rules-and-evidence layer between the rail and the runtime.

Governance latency (sub-second invariant)

"Sub-second governance" is a tested invariant, not marketing. The bench in tests/bench/governance-latency.bench.ts measures each governance hot path with vitest bench and emits a grep-able [gov-bench] summary line per scenario. Numbers below are steady-state percentiles from npm run bench:governance (run on the dev machine — your hardware will differ; the relative ordering is what matters).

Both per-event hot paths (evaluateAction, MerkleAudit.record) clear an entire EU AI Act policy gate and audit-chain write two orders of magnitude inside one millisecond. The end-to-end remember() path — including Ed25519 sign + sequence + chain emit — still sits comfortably inside the "sub-second governance" envelope with three decimal-orders of headroom.

A CI-enforced guard spec in tests/governance/latency-invariant.test.ts runs a degraded-mode sample on every npm test and fails if p95 for policy.evaluateAction regresses past 1 ms, or MerkleAudit.record past 5 ms. Bounds are sized to catch a ~10x regression, not flap on jitter.

How to reproduce: npm run bench:governance (full vitest.bench harness) or npm test -- latency-invariant (CI-bound check).

Native rails

Every rail ships with the same PaymentRail interface as StripeRail / PaystackRail / LightningRail:

Plus the Spatial governance fold — attachSpatialEvidence() co-signs the MerkleAudit chain with GridStamp proof-of-presence for embodied agents (drones, robots). Loose-coupled — no gridstamp runtime dependency.

Subpath imports for smaller, safer consumers

If you only need one MnemoPay module, import that subpath instead of the package root. This keeps MCP servers and other stdio tools quiet, avoids pulling unused middleware into bundles, and makes the dependency boundary obvious.

import { localEmbed, cosineSimilarity } from "@mnemopay/sdk/recall";
import { StripeRail, X402Rail } from "@mnemopay/sdk/rails";
import { SQLiteStorage } from "@mnemopay/sdk/storage";
import { CommerceEngine } from "@mnemopay/sdk/commerce";

Use the root import when you want the full SDK surface. Use @mnemopay/sdk/mcp only when you are intentionally mounting the MnemoPay MCP server.

Swarm (stable — v1.11+)

@mnemopay/sdk/swarm is the missing piece that browse.sh shipped as a public skill catalog. Ours adds the bit they don't: every agent in the swarm carries a DID, every action is FiscalGate-prechecked against per-agent + total caps, every TaskResult is appended to a shared Article-12 audit chain, and every skill invocation is billable through the same hash-chained ledger the rest of the SDK already uses.

CLI: npx @mnemopay/swarm list · npx @mnemopay/swarm demo — see mnemopay-swarm.

import { Swarm } from "@mnemopay/sdk/swarm";
import { AuditChain } from "@mnemopay/sdk/governance";
import { open } from "@mnemopay/browser";   // any BrowserProvider works

const provider = await someProviderFactory();
const swarm = new Swarm({
  size: 4,
  provider,
  did: "did:mp:abc...",
  budget: { perAgent: 0.25, total: 1.00 },
  audit: { chain: new AuditChain() },
});

const run = await swarm.spawn([
  { id: "t1", skillId: "ramp.com/expense-create",  prompt: "submit $42 lunch" },
  { id: "t2", skillId: "linear/issue-create",      prompt: "file UI bug" },
  { id: "t3", skillId: "cloudflare/dns-record-set", prompt: "add CNAME" },
]);

const results = await swarm.gather(run);
const final   = await swarm.recombine(results, "merge-json");

When to use it. Any time you'd open N parallel browser sessions to attack a problem — multi-source research, cross-platform issue triage, A/B-style "ask three agents, take the majority answer" — but you want one audit bundle, one budget envelope, and one place where billing happens.

Three recombine strategies (plus your own).

• first-success — returns the output of the first ok:true task; perfect for race patterns where any answer is fine.
• majority-vote — returns the most-common output across ok:true tasks; perfect for fact-extraction where consensus matters.
• merge-json — deep-merges every ok:true object output with sorted keys (deterministic across runs).
• concat — joins string outputs with \n in spawn order.
• Or pass any (results) => unknown callback.

Skill catalog. Public listings live at mcp.mnemopay.com/skills. The catalog is intentionally small and honestly marked — verified-partner badges only show after a real partnership is signed. Everything else carries verified: false, status: 'pending-partner' so you know exactly what trust tier you're getting.

BrowserSwarm — native browser-session fan-out (stable since 1.11.0)

@mnemopay/sdk/swarm/browser extends Swarm with a typed step sequence (goto / act / extract / screenshot / wait) per task and a lazy wire to @mnemopay/browser (optional peer dep — installing the SDK does NOT pull Playwright). Each task gets its own browser session, every step appends a browser.step event to the shared audit chain, and a thrown step kills only that one task — sibling sessions keep running.

import { BrowserSwarm } from "@mnemopay/sdk/swarm/browser";
const swarm = new BrowserSwarm({
  size: 3, provider: undefined as never,
  budget: { perAgent: 0.25, total: 1.00 },
  browser: { provider: "stagehand" },
});
const run = await swarm.spawn([
  { id: "amzn", prompt: "amazon price",  steps: [{type:"goto", url:"https://amazon.com/dp/X"},  {type:"extract", selector:"#priceblock_ourprice"}] },
  { id: "bby",  prompt: "best buy price", steps: [{type:"goto", url:"https://bestbuy.com/site/X"},{type:"extract", selector:".priceView-customer-price"}] },
  { id: "tgt",  prompt: "target price",   steps: [{type:"goto", url:"https://target.com/p/X"},   {type:"extract", selector:"[data-test=product-price]"}] },
]);
const results = await swarm.gather(run);   // BrowserTaskResult[] with .screenshots + .extractedData

File issues at github.com/mnemopay/mnemopay-sdk.

Audit-only middleware — .audit(client) with streaming + on-disk chain (1.11.0-alpha.0)

For chat widgets and regulated pipelines where ANY prompt mutation is a violation but Article-12 telemetry is still required:

import { AuditChain } from "@mnemopay/sdk/governance/audit-chain";
import { AnthropicMiddleware } from "@mnemopay/sdk/middleware/anthropic-audit";

const chain = new AuditChain({ path: "./.audit-chain/llm.jsonl" });  // file-backed since 1.11.0-alpha.0
const client = AnthropicMiddleware.audit(new Anthropic(), { chain });

// .create AND .stream now both emit one `llm.call` event per call. Streams
// that get cancelled mid-iteration emit `partial: true` with tokens-so-far.
for await (const chunk of client.messages.stream({ model, max_tokens, messages })) { /* ... */ }

@mnemopay/sdk/middleware/openai-audit exposes the equivalent shape for OpenAI — chat.completions.create({ stream: true }) is intercepted automatically (pass stream_options: { include_usage: true } to capture the final usage block).

Building an MCP server? Start here.

If you're shipping an MCP server and want to charge per-call — even sub-cent amounts — MnemoPay is built for you.

• Sub-cent payments via Lightning rail (impossible on Stripe/Paystack due to fees)
• Per-tool metering with agent.charge(amount, toolName) — two lines of code
• Agent Reputation Scoring gates abusive callers automatically — 300-850 reputation score, free tier + paid tier
• Cryptographic receipts every user can audit — no "trust me bro" billing
• Free indefinitely for the first 10 MCP servers that adopt it, subject to 90 days' written notice of any future change (email with your repo)

import MnemoPay from "@mnemopay/sdk";
const agent = MnemoPay.quick("my-mcp-server");

// Inside your tool handler:
const tx = await agent.charge(0.002, "embed_document");  // 0.2¢
if (tx.status === "blocked") return { error: "Payment declined" };
await agent.settle(tx.id);
// ... run the tool

Zero-config starter → production Lightning rail → Agent Reputation Scoring gating. Same API.

What Makes MnemoPay Different

$87M has been invested across 5 competitors. None have more than 3 of these 10 features:

Agent Reputation Scoring

A novel cross-session reputation scoring system for AI agents. Five-component scoring on a 300-850 range (familiar to developers from consumer credit; MnemoPay is not affiliated with Fair Isaac Corporation or any consumer credit bureau):

import { AgentReputationScoring } from "@mnemopay/sdk";

const scorer = new AgentReputationScoring();
const result = scorer.compute({
  transactions: await agent.history(1000),
  createdAt: agentCreationDate,
  fraudFlags: 0,
  disputeCount: 0,
  disputesLost: 0,
  warnings: 0,
  budgetCap: 5000,
  memoriesCount: agent.memories.size,
});

console.log(result.score);      // 742
console.log(result.rating);     // "very_good"
console.log(result.feeRate);    // 0.013 (1.3%)
console.log(result.trustLevel); // "high"
console.log(result.requiresHITL); // false

Behavioral Finance Engine

Peer-reviewed behavioral economics from Nobel laureate Daniel Kahneman and collaborators. Every parameter cited to published research.

import { BehavioralEngine } from "@mnemopay/sdk";

const behavioral = new BehavioralEngine();

// Prospect Theory (Kahneman & Tversky, 1992)
// Losses hurt 2.25x more than gains feel good
behavioral.prospectValue(100);   // { value: 57.5, domain: "gain" }
behavioral.prospectValue(-100);  // { value: -129.5, domain: "loss" }

// Should the agent wait before buying?
const cooling = behavioral.coolingOff(2000, 5000); // amount, monthly income
// → { recommended: true, hours: 3.2, riskLevel: "high", regretProbability: 0.65 }

// Frame spending as goal delay (2.25x more effective than gain framing)
const frame = behavioral.lossFrame(200, {
  name: "Emergency Fund", target: 10000, current: 3000, monthlySavings: 500
});
// → "This $200 purchase delays your Emergency Fund goal by 12 days."

// Save More Tomorrow (Thaler & Benartzi, 2004)
const smart = behavioral.commitmentDevice(0.035, 0.03, 4);
// → { finalRate: 0.095, explanation: "3.5% → 9.5% over 4 raise cycles" }

// Predict regret from purchase history
behavioral.recordRegret({ amount: 300, category: "gadgets", regretScore: 8, timestamp: "..." });
const prediction = behavioral.predictRegret(400, "gadgets");
// → { probability: 0.72, triggerCoolingOff: true }

Research sources: Tversky & Kahneman 1992, Laibson 1997, Thaler & Benartzi 2004, Barber & Odean 2000, Nunes & Dreze 2006, Shiller 2000.

Memory Integrity (Merkle Tree)

Tamper-evident memory. If anyone injects, modifies, or deletes an agent's memories, the Merkle root changes and you know.

import { MerkleTree } from "@mnemopay/sdk";

const tree = new MerkleTree();

// Every memory write adds a leaf
tree.addLeaf("mem-1", "User prefers monthly billing");
tree.addLeaf("mem-2", "Last purchase was $25 API access");

// Take periodic snapshots
const snapshot = tree.snapshot();
// → { rootHash: "a3f2...", leafCount: 2, snapshotHash: "b7c1..." }

// Later: check if memories were tampered
const check = tree.detectTampering(snapshot);
// → { tampered: false, summary: "Integrity verified. 2 memories, root matches." }

// Prove a specific memory exists without revealing others
const proof = tree.getProof("mem-1");
MerkleTree.verifyProof(proof); // true

Defends against: MemoryGraft injection, silent deletion, content tampering, replay attacks, reordering attacks.

Anomaly Detection (EWMA + Behavioral Fingerprinting + Canaries)

Three independent systems that catch compromised agents.

import { EWMADetector, BehaviorMonitor, CanarySystem } from "@mnemopay/sdk";

// 1. EWMA: real-time streaming anomaly detection
const detector = new EWMADetector(0.15, 2.5, 3.5, 10);
detector.update(100); // normal
detector.update(100); // normal
detector.update(9999); // → { anomaly: true, severity: "critical", zScore: 8.2 }

// 2. Behavioral fingerprinting: detect hijacked agents
const monitor = new BehaviorMonitor({ warmupPeriod: 10 });
// Build profile over time
monitor.observe("agent-1", { amount: 100, hourOfDay: 14, chargesPerHour: 2 });
// Sudden change = suspected hijack
monitor.observe("agent-1", { amount: 9999, hourOfDay: 3, chargesPerHour: 50 });
// → { suspected: true, severity: "critical", anomalousFeatures: 3 }

// 3. Canary honeypots: plant traps for compromised agents
const canary = new CanarySystem();
const trap = canary.plant("transaction");
canary.check(trap.id, "rogue-agent");
// → { severity: "critical", message: "CANARY TRIGGERED: Agent compromised" }

Math: mu_t = alpha * x_t + (1 - alpha) * mu_{t-1}, alert when |x_t - mu_t| > k * sigma_t (Roberts 1959, Lucas & Saccucci 1990).

Memory (Compounding Knowledge Base)

Not a traditional RAG lookup. MnemoPay memories compound — every transaction strengthens associated context, weak memories decay, strong ones consolidate. The same pattern Karpathy describes as "LLM Wiki" but applied to payments and trust.

• Ebbinghaus forgetting curve — memories decay naturally over time
• Hebbian reinforcement — successful transactions strengthen associated memories
• RL feedback loop — rlFeedback(ids, reward) applies EWMA importance updates after agent actions
• Consolidation — auto-prunes weak memories, keeps what matters
• Semantic recall — find memories by relevance, not just recency
• 100KB per memory — store rich context, not just strings

// After a recall + action, signal usefulness with rlFeedback
const memories = await agent.recall("user preferences", 5);
// ... agent acts on recalled memories ...
await agent.rlFeedback(memories.map(m => m.id), +1.0);   // +1 = useful, -1 = not useful

Choosing a persistence adapter

Recall is backed by a pluggable PersistenceAdapter. Pick by deployment shape:

PostgresAdapter and NeonAdapter are the same pgvector-backed implementation — "Neon" is just hosted Postgres; use whichever name fits your infra.

import { MnemoPay } from "@mnemopay/sdk";

// Via MnemoPay.create — { type: "postgres" } (alias of "neon")
const agent = await MnemoPay.create({
  agentId: "agent-1",
  persist: { type: "postgres", url: process.env.DATABASE_URL! },
});

// Or construct the adapter directly
import { PostgresAdapter, postgresMigrationSql } from "@mnemopay/sdk/recall/postgres";
const adapter = new PostgresAdapter({ url: process.env.DATABASE_URL! });

The schema (a vector(384) column + HNSW cosine index) is auto-created on the first write. To manage it with your own migration tool instead, run the DDL from postgresMigrationSql(table?, dimensions?) and pass skipBootstrap: true. Requires the optional peer dep: npm install pg.

Reputation Streaks & Badges

Agents earn trust over time. Consecutive successful settlements build streaks that unlock badges and reduce fees.

const rep = await agent.reputation();
console.log(rep.streak);
// → { currentStreak: 47, bestStreak: 312, streakBonus: 0.094 }

console.log(rep.badges);
// → [
//   { id: "first_settlement", name: "First Settlement", earnedAt: 1712700000000 },
//   { id: "streak_50", name: "Streak Master", earnedAt: 1712900000000 },
//   { id: "volume_10k", name: "High Roller", earnedAt: 1713100000000 },
// ]

Streaks reset on refunds or disputes. Streak bonuses compound reputation up to +10%.

Hash-Chained Ledger

Every ledger entry links to the previous via SHA-256 hash chain. If any entry is modified, the chain breaks and verify() catches it instantly.

const summary = agent.ledger.verify();
console.log(summary.chainValid);     // true
console.log(summary.chainIntegrity); // 1.0 (100% of links verified)

Combined with Merkle integrity on memories and HMAC on transactions, MnemoPay gives you three independent tamper-detection systems.

Payments (cent-precise double-entry)

• Double-entry bookkeeping — every debit has a credit, always balances to zero
• Escrow flow — charge -> hold -> settle -> refund (same shape as Stripe/Square)
• Volume-tiered fees — 1.9% / 1.5% / 1.0% based on cumulative volume
• 3 payment rails — Paystack (Africa), Stripe (global), Lightning (BTC)
• Cent-precise integer math — stress-tested with 200,000 transactions across 50 concurrent agents, zero drift

Identity (KYA Compliance)

• Cryptographic identity — HMAC-SHA256 keypairs, replay protection
• Capability tokens — scoped permissions with spend limits
• Counterparty whitelists — restrict who the agent can transact with
• Kill switch — revoke all tokens instantly

Fraud Detection (ML-grade)

• Velocity checks — per-minute/hour/day limits
• Isolation Forest — unsupervised ML anomaly detection
• Geo-enhanced — country tracking, rapid-hop detection, OFAC sanctions
• Adaptive engine — asymmetric AIMD, anti-gaming, circuit breaker, PSI drift detection

Multi-Agent Commerce

• CommerceEngine — autonomous shopping with mandates, escrow, approval callbacks
• MnemoPayNetwork — register agents, execute deals, shared memory context
• Supply chains — 10-step agent chains, 100-agent marketplaces, all tested

Claude Agent SDK integration

Two primitives built specifically for the Claude Agent SDK pattern where an Opus orchestrator spawns Sonnet/Haiku subagents.

1-hour prompt cache on recall results

When you feed MnemoPay recall into a Claude system prompt, use formatForClaudeCache() to emit a content block with cache_control: { type: "ephemeral", ttl: 3600 }. The Anthropic API caches that prefix for up to 1 hour; cache reads are billed at roughly 10% of the normal input rate. With stable recall prefixes and a warm 1h cache, users have observed savings in the typical range of 85-92% on the recall portion of input tokens — your actual results depend on call frequency and memory set stability.

import MnemoPay, { formatForClaudeCache } from "@mnemopay/sdk";
import Anthropic from "@anthropic-ai/sdk";

const agent = MnemoPay.quick("my-agent");
const anthropic = new Anthropic();

// Option A: recall() directly returns a cache block
const cacheBlock = await agent.recall("user preferences", 10, {
  formatForClaudeCache: true,
});

// Option B: convert an existing memory array (no extra recall call)
const memories = await agent.recall("user preferences", 10);
const cacheBlock2 = MnemoPay.formatForClaudeCache(memories);
// OR: formatForClaudeCache(memories) from the module directly

const response = await anthropic.messages.create({
  model: "claude-opus-4-7",
  max_tokens: 1024,
  system: [
    { type: "text", text: "You are a helpful assistant.", cache_control: { type: "ephemeral" } },
    cacheBlock,  // ← MnemoPay recall cached for 1 hour
  ],
  messages: [{ role: "user", content: userMessage }],
});

The serialized text is sorted by memory id so identical memory sets produce byte-identical output — required for the cache prefix to hit on subsequent turns.

Per-subagent cost attribution

Track which subagent in a multi-agent pipeline spent how much — recorded as double-entry ledger pairs so it stays audit-clean.

import MnemoPay, { SubagentCostTracker } from "@mnemopay/sdk";

const orchestrator = MnemoPay.quick("orchestrator");

// After each Claude API call, record the cost:
orchestrator.subagentCosts.attributeSubagentCost({
  parentAgentId: "orchestrator",
  subagentId: "researcher-1",
  subagentRole: "researcher",
  modelId: "claude-sonnet-4-6",
  inputTokens: 5000,
  outputTokens: 2000,
  cacheReadTokens: 8500,   // tokens served from the 1h recall cache
  cacheWriteTokens: 500,
  cacheWriteTtl: "1h",
});

// At end of pipeline, get breakdown ordered by cost:
const breakdown = orchestrator.subagentCosts.subagentCostBreakdown("orchestrator");
// → [{ subagentId, subagentRole, modelId, totalCostUsd, cacheSavingsUsd, ... }]

const totalSaved = orchestrator.subagentCosts.totalCacheSavings("orchestrator");

Pricing table used: 2026 Anthropic list rates (Opus 4.7 $5/$25/M, Sonnet 4.6 $3/$15/M, Haiku 4.5 $1/$5/M; cache reads 0.1×, 1h writes 2×). Update MODEL_PRICING in src/subagent-cost.ts if rates change.

See docs/agent-sdk-guide.md for a full integration walkthrough.

Payment Rails

Every rail implements the same PaymentRail interface — createHold / capturePayment / reversePayment. Swap rails without touching agent code.

import {
  PaystackRail, StripeRail, LightningRail,
  StripeMPPRail, X402Rail, GoogleAP2Rail,
} from "@mnemopay/sdk";

const paystack  = new PaystackRail(process.env.PAYSTACK_SECRET_KEY!);
const stripe    = new StripeRail(process.env.STRIPE_SECRET_KEY!);
const lightning = new LightningRail(LND_URL, MACAROON);

const mpp   = new StripeMPPRail(process.env.STRIPE_SECRET_KEY!);
const x402  = new X402Rail({ signer: yourEip3009Signer });   // bring-your-own crypto
const ap2   = new GoogleAP2Rail({ mandate, endpoint, signer });

const agent = MnemoPay.quick("my-agent", { paymentRail: paystack });

Stripe — real card charges with saved customers

End-to-end flow for charging a user's saved card without a browser handoff:

import MnemoPay, { StripeRail } from "@mnemopay/sdk";

const rail = new StripeRail(process.env.STRIPE_SECRET_KEY!);
const agent = MnemoPay.quick("agent-1", { paymentRail: rail });

// 1. Create a Stripe customer (one-time, persist cus_... to your DB)
const { customerId } = await rail.createCustomer("user@example.com", "Jerry O");

// 2. Collect a card via Stripe.js: create a SetupIntent, return client_secret
//    to the browser, let Stripe Elements confirm it. You receive pm_... from
//    the webhook or confirmation callback. Save it alongside the customer.
const { clientSecret } = await rail.createSetupIntent(customerId);
// → hand clientSecret to frontend, get back paymentMethodId after confirm

// 3. Charge the saved card later, off-session, no user interaction needed
const tx = await agent.charge(25, "Monthly API access", undefined, {
  customerId,
  paymentMethodId: "pm_saved_from_step_2",
  offSession: true,
});

// 4. Settle (captures the hold) or refund (releases it)
await agent.settle(tx.id);

Paystack supports the same pattern via authorizationCode:

const tx = await agent.charge(5000, "NGN invoice", undefined, {
  email: "customer@example.com",
  authorizationCode: "AUTH_abc123", // from an earlier Paystack transaction
});

MCP Server

npx @mnemopay/sdk init
# or
claude mcp add mnemopay -s user -- npx -y @mnemopay/sdk

Default tool group: essentials (14 tools, ~1K tokens). One of the lightest MCP servers you can install — MnemoPay only loads memory + wallet + tx by default so it doesn't tax your agent's context budget.

• memory: remember, recall, forget, reinforce, consolidate
• wallet: balance, profile, history, logs
• tx: charge, settle, refund, dispute, receipt_get

Need more? Opt in explicitly:

npx @mnemopay/sdk --tools=all       # all 95 tools
npx @mnemopay/sdk --tools=agent     # essentials + commerce + hitl + payments + webhooks
npx @mnemopay/sdk --tools=reputation  # Agent Reputation Scoring only

Groups: memory, wallet, tx, commerce, hitl, payments, webhooks, reputation, security, governance, identity, skills, spatial, agent_os, organization_admin, operator. Aliases: essentials (default), agent, all. Also settable via MNEMOPAY_TOOLS env var.

Breaking change in v1.3.0: default was all, now essentials. If you relied on commerce/hitl/webhooks/fico/security being available without a flag, pass --tools=all or --tools=agent. See CHANGELOG.

Middleware

Drop-in proxies that make recall invisible: every chat call auto-injects the top memories as system context and stores the exchange afterward. Same Middleware.wrap(client, agent) shape across every provider.

// OpenAI
import { mnemoPayMiddleware } from "@mnemopay/sdk/middleware/openai";

// Anthropic
import { mnemoPayMiddleware } from "@mnemopay/sdk/middleware/anthropic";

// Gemini
import { GeminiMiddleware } from "@mnemopay/sdk/middleware/gemini";

// Cohere (v2 chat API)
import { CohereMiddleware } from "@mnemopay/sdk/middleware/cohere";
const cohere = CohereMiddleware.wrap(new CohereClientV2({ token }), agent);

// Mistral
import { MistralMiddleware } from "@mnemopay/sdk/middleware/mistral";
const mistral = MistralMiddleware.wrap(new Mistral({ apiKey }), agent);

// LangGraph
import { mnemoPayTools } from "@mnemopay/sdk/langgraph";

Architecture

Full stack diagram and module map: docs/architecture.md.

┌──────────────────────────────────────────────────────────────────┐
│                       MnemoPay SDK                                │
│              Governance · Memory · Payments · Identity            │
├─────────────────────────────────────────────────────────────────┤
│ GOVERNANCE  Charter · FiscalGate · Article 12 · MerkleAudit      │
│             mission scope, budget enforcement, audit bundles     │
├──────────┬──────────┬───────────┬─────────────────────────────────┤
│  Memory  │ Payments │ Identity  │  Agent Reputation Scoring       │
│          │          │           │  300-850, 5-component           │
│ remember │ charge   │ KYA       ├─────────────────────────────────┤
│ recall   │ settle   │ tokens    │  Behavioral Finance             │
│ reinforce│ refund   │ perms     │  prospect theory, nudges        │
│ forget   │ dispute  │ killswitch├─────────────────────────────────┤
│          │          │           │  Anomaly Detection              │
│          │          │           │  EWMA + fingerprinting          │
├──────────┴──────────┴───────────┼─────────────────────────────────┤
│     Double-Entry Ledger         │  Merkle Integrity               │
│  debit + credit = always zero   │  tamper-evident memory          │
├─────────────────────────────────┼─────────────────────────────────┤
│     Fraud Guard (ML-grade)      │  Canary Honeypots               │
│  velocity + geo + adaptive      │  compromise detection           │
├─────────────────────────────────┴─────────────────────────────────┤
│ SPATIAL    GridStamp adapter — proof-of-presence for embodied      │
│            agents (drones, robots). Loose-coupled, fail-closed.    │
├──────────────────────────────────────────────────────────────────┤
│ RAILS  Stripe · Paystack · Lightning · StripeMPP · x402 · AP2    │
│        same PaymentRail interface — drop-in swap, no agent diff   │
└──────────────────────────────────────────────────────────────────┘

Module stability

MnemoPay follows semver. Stability tiers tell you how much a module's public API may shift before 2.0 — see VERSIONING.md for the full contract.

Testing

npm test    # full test suite across 12 files

• core.test.ts — memory, payments, lifecycle, reputation scoring, behavioral, Merkle, EWMA, canaries, streaks, badges
• fraud.test.ts — velocity, anomaly, fees, disputes, replay detection
• geo-fraud.test.ts — geo signals, trust, sanctions
• identity.test.ts — KYA, tokens, permissions
• production-100k.test.ts — 100K operations, 10 concurrent agents, hash-chain verification, zero drift
• stress-200k.test.ts — 200K real-world stress: 50 agents, burst traffic, race conditions, refund storms, memory leak detection
• ledger.test.ts — double-entry, reconciliation
• network.test.ts — multi-agent, deals, supply chains
• paystack.test.ts — rail, webhooks, transfers
• stress.test.ts — 1000-cycle precision, parallel ops
• recall.test.ts — semantic search, decay, reinforcement

License

Apache License 2.0 — see LICENSE.

Copyright 2026 J&B Enterprise LLC.

Third-party attributions

The entity-observation write-path in src/recall/observations.ts (per-entity consolidated summaries, debounced regeneration, session-spanning rollups) is derived from vectorize-io/hindsight (MIT, Copyright (c) 2025 Vectorize AI, Inc.). The full upstream notice is preserved in NOTICE and in the header of the ported file.

Trademark and regulatory notices

Agent Reputation Scoring is a trustworthiness scoring system for autonomous software agents, not for consumer credit reporting. It does not produce a consumer report as defined by the Fair Credit Reporting Act (FCRA) and is not regulated under the FCRA. MnemoPay is not a consumer reporting agency.

MnemoPay is not a bank, money transmitter, or insurer, and does not hold customer deposits. Payments are settled through third-party payment rails (Stripe, Paystack, Lightning Network) — MnemoPay is software that connects to those rails on behalf of developers, not a financial institution.

"FICO" is a registered trademark of Fair Isaac Corporation. MnemoPay and its Agent Reputation Scoring module are not affiliated with, endorsed by, or derived from Fair Isaac Corporation. The AgentCreditScore and AgentFICO export names are deprecated aliases kept for backward compatibility with earlier beta releases and will be removed in a future major version.

Built by Jeremiah Omiagbo