mnemopay 1.2.1

Memory + Wallet for AI Agents. Give any agent persistent memory and micropayment capabilities.

MnemoPay Python SDK

The governance layer for AI agents that handle money — Python edition. Mirrors the TypeScript @mnemopay/sdk shape so the same agent identity, score, charter, and audit chain work across both runtimes.

from mnemopay import MnemoPay

agent = MnemoPay("my-agent")
agent.remember("user prefers Python")
tx = agent.charge(10.00, "API call")
# Settle after the escrow hold window expires.
agent.settle(tx.id)

Install

pip install mnemopay                  # stdlib-only core
pip install "mnemopay[stripe]"        # + StripeRail (peer-loaded `stripe>=12.0`)

What MnemoPay is NOT: not a bank, not a money transmitter, not a Stripe replacement, not an agent framework. It sits above the rail and below the runtime — declares the rules, enforces the budget, produces the evidence.

Features

• Payment rails (TS parity): MockRail, StripeRail, PaystackRail, LightningRail — manual-capture two-phase commit, threading-safe capture race-protection, idempotency-key forwarding
• Cognitive Memory: Ebbinghaus decay, Hebbian reinforcement, auto-scoring, Layer 2 semantic recall + RL feedback
• Micropayments: escrow-based charges with volume-tiered fees (1.9% / 1.5% / 1.0%)
• Agent Credit Score: FICO-style agent behavioral scoring (300-850 range; not consumer FICO, not FCRA-regulated, not affiliated with Fair Isaac Corporation). Importable as AgentCreditScore (preferred) or the deprecated AgentFICO alias.
• Behavioral Finance: prospect theory, cooling-off periods, regret prediction
• Merkle Integrity: SHA-256 tamper detection for agent memory
• Anomaly Detection: EWMA streaming detector, behavioral fingerprinting, canary honeypots
• Commerce: autonomous shopping engine with mandates + escrow
• Circuit Breaker: AIMD rate limiting, anti-gaming, PSI drift detection
• Stdlib core: zero required dependencies — peer-deps only when you opt in (e.g. stripe)
• Python 3.9+: full type hints, dataclasses, sync API (matches the rest of the SDK)

Modules

Module	Class	Description
mnemopay.core	MnemoPay	Memory + payments + reputation
mnemopay.rails	PaymentRail, MockRail, StripeRail, PaystackRail, LightningRail	Payment rail abstraction
mnemopay.agent_credit_score	AgentCreditScore	Credit scoring (300-850, FICO-style; not affiliated with Fair Isaac Corporation). The mnemopay.fico module + AgentFICO class remain as deprecated aliases.
mnemopay.behavioral	BehavioralEngine	Behavioral finance tools
mnemopay.integrity	MerkleTree	SHA-256 tamper detection
mnemopay.anomaly	EWMADetector, BehaviorMonitor, CanarySystem	Anomaly detection
mnemopay.commerce	CommerceEngine, CommerceProvider, Mandate	Autonomous shopping
mnemopay.circuit_breaker	CircuitBreaker, AIMDRateLimiter, AntiGamingEngine, PSIDriftDetector	Adaptive defense

Payment rails

Mirrors the TypeScript PaymentRail interface. Same shape (create_hold / capture_payment / reverse_payment), same drop-in-swap semantics. Sync API to match the rest of the Python SDK — no asyncio.

from mnemopay.rails import StripeRail, MockRail, HoldOptions

# Default — no infra, in-memory ledger, used in tests + dev
rail = MockRail()

# Production — real Stripe PaymentIntents with manual-capture two-phase commit
rail = StripeRail(secret_key="sk_test_...")

# Two-phase commit (hold → capture)
hold = rail.create_hold(
    amount=25.00,
    reason="Monthly access",
    agent_id="agent-1",
    opts=HoldOptions(
        customer_id="cus_real",
        payment_method_id="pm_real",
        off_session=True,
        metadata={"idempotencyKey": "req_abc"},
    ),
)
capture = rail.capture_payment(hold.external_id, 25.00)
# capture.status == "succeeded"
# capture.receipt_id == "ch_..."  # Stripe charge id

# Or reverse (cancel the hold) instead of capturing
rail.reverse_payment(hold.external_id, 25.00)

StripeRail includes onboarding helpers for off-session charges:

result = rail.create_customer("[e-mail]", name="Ada O")
# {"customer_id": "cus_..."}

result = rail.create_setup_intent(customer_id="cus_...")
# {"setup_intent_id": "seti_...", "client_secret": "..."}

For tests, inject a mock client:

from unittest.mock import MagicMock
rail = StripeRail.from_client(MagicMock(), currency="usd")

Compatibility with @mnemopay/sdk (TypeScript)

Feature	TypeScript v1.12.0	Python v1.1.1
MockRail	yes	yes
StripeRail	yes	yes
PaystackRail	yes	yes
LightningRail	yes	yes
StripeMPPRail (alpha)	yes	not yet
X402Rail (alpha)	yes	not yet
GoogleAP2Rail (alpha)	yes	not yet
Governance (charter / FiscalGate / Article 12 / risk taxonomy / action ledger / governed skills)	yes	not yet
Agent Credit Score	yes	yes
Behavioral Finance	yes	yes
Merkle integrity	yes	yes
Anomaly detection (EWMA + canary)	yes	yes
Commerce engine	yes	yes
Circuit breaker	yes	yes

The Python SDK ships behind the TypeScript SDK — port priorities track agent-developer demand. Open an issue if you need a specific rail or governance primitive in Python.

Tests

pip install -e ".[dev,stripe]"
pytest                    # 435 tests across 9 modules
pytest tests/test_rails.py -v   # 29 rail-specific tests

License

v1.0.0+ is Apache License 2.0 — see LICENSE and NOTICE.

The pre-release betas (1.0.0b1 through 1.0.0b4) shipped under the MIT License and remain available under those terms on PyPI in perpetuity. If you depend on the MIT terms specifically, pin to mnemopay==1.0.0b4.

Why Apache 2.0 at the stable cut: alignment with the canonical TypeScript SDK at @mnemopay/sdk plus the patent grant + retaliation clause for enterprise embedders.

Copyright 2026 J&B Enterprise LLC.

Trademark notice

FICO is a registered trademark of Fair Isaac Corporation. MnemoPay's Agent Credit Score is FICO-style behavioral scoring (300-850 range, five-component methodology) and is not affiliated with or endorsed by Fair Isaac Corporation. Consumer FICO scores are regulated under the FCRA and produced by Fair Isaac Corporation; this SDK produces neither.