Lightning Memory

Lightning Memory

Persistent memory for AI agents in the Lightning economy.

The Problem

AI agents spend sats over Lightning via L402 — but they can't remember what they bought. Every session starts from zero. Every vendor is a stranger. Every price is accepted at face value. An agent that paid 500 sats yesterday doesn't know if today's 5,000 sat invoice is a price spike or normal.

The Solution

L1: Bitcoin      — settles

L2: Lightning — pays L3: Lightning Memory — remembers

Lightning Memory gives agents persistent memory, vendor intelligence, and payment safety gates. Agents learn from their spending history, track vendor reputations, detect price anomalies, enforce budgets, and share trust signals with other agents.

[Interactive Demo](https://www.jasonsosa.com/blog/agent-lightning-memory) — watch an agent learn, get rugged, and route around bad actors.

[Building the Agent Economy](https://www.jasonsosa.com/blog/agent-economy-trust-marketplace-bitcoin-lightning) — trust, budgets, compliance, and the memory marketplace.

Who Is This For

• Agents making L402 payments that need vendor reputation and spending discipline
• Developers building autonomous agents on Bitcoin/Lightning
• Anyone running an MCP-compatible AI agent (Claude, GPT, or any MCP client)

Quick Start

pip install lightning-memory

lightning-memory # starts MCP server

Configure in Claude Code

{

"mcpServers": { "lightning-memory": { "command": "lightning-memory" } } }

Configure in Claude Desktop

{

"mcpServers": { "lightning-memory": { "command": "python", "args": ["-m", "lightning_memory.server"] } } }

How It Compares

Tools (22)

Memory

memory_store(

content="Paid 500 sats to bitrefill.com for a $5 Amazon gift card via L402.", type="transaction", metadata='{"vendor": "bitrefill.com", "amount_sats": 500}' )

memory_query(query="bitrefill payment history", limit=5)

→ recency-weighted results with dedup and contradiction alerts

Lightning Intelligence

ln_vendor_reputation(vendor="bitrefill.com")

→ {reputation: {total_txns: 12, success_rate: 0.92, avg_sats: 450}, recommendation: "reliable"}

ln_anomaly_check(vendor="bitrefill.com", amount_sats=5000)

→ {anomaly: {verdict: "high", context: "5000 sats is 11.1x the historical average..."}}

Payment Safety

ln_preflight(vendor="bitrefill.com", amount_sats=500)

→ {decision: {verdict: "approve", budget_remaining_today: 4500, trust_score: 0.89}}

If the vendor suddenly charges 50x:

ln_preflight(vendor="bitrefill.com", amount_sats=25000)

→ {decision: {verdict: "reject", reasons: ["exceeds daily limit of 5000 sats"]}}

Trust & Compliance

Marketplace

ln_discover_gateways(operation="memory_query")

→ {count: 2, gateways: [{url: "https://gw1.example.com", operations: {...}}, ...]}

ln_remote_query( gateway_url="https://gw.example.com", operation="ln_vendor_reputation", params='{"vendor": "openai"}' )

→ Pays 3 sats, returns remote agent's vendor intelligence

Architecture

• Nostr identity: Agent identity = Nostr keypair (BIP-340). No accounts, no API keys.
• Local-first: SQLite with FTS5 full-text search + optional ONNX semantic search. Works offline.
• Nostr sync: Memories written as NIP-78 events to relays. Portable, tamper-proof.
• L402 payments: Pay-per-query gateway. 1-10 sats per operation.
• Memory quality: Deduplication, contradiction detection, noise filtering, recency-weighted ranking, access tracking.

L402 Gateway

Run an L402 pay-per-query HTTP gateway. Other agents pay Lightning micropayments to access your agent's memory — no API keys, no accounts.

pip install lightning-memory[gateway]

lightning-memory-gateway # Listening on 0.0.0.0:8402

How L402 Works

Agent                          Gateway                      Phoenixd

Endpoints

Phoenixd Setup

1. Download and run Phoenixd (listens on localhost:9740)
2. Fund it with ~10,000 sats for initial channel opening
3. Configure: ~/.lightning-memory/config.json → {"phoenixd_password": " "}
4. Start: lightning-memory-gateway

Docker

PHOENIXD_PASSWORD=your-password docker compose up

CLI Commands

lightning-memory                  # Start MCP server

lightning-memory stats # Memory statistics dashboard lightning-memory export json # Export memories as JSON lightning-memory export csv # Export memories as CSV lightning-memory relay-status # Check Nostr relay connectivity lightning-memory-gateway # Start L402 HTTP gateway lightning-memory-manifest # Generate gateway discovery manifest

Relay Configuration

Default relays: wss://relay.damus.io, wss://nos.lol, wss://relay.nostr.band

Customize in ~/.lightning-memory/config.json:

{

"relays": ["wss://relay.damus.io", "wss://nos.lol", "wss://relay.primal.net"], "sync_timeout_seconds": 30, "max_events_per_sync": 500 }

Optional: Semantic Search

Add ONNX-based semantic similarity search alongside FTS5 keyword search:

pip install lightning-memory[semantic]

Queries then use hybrid ranking: FTS5 BM25 + cosine similarity with reciprocal rank fusion. "Which vendors are reliable for transcription" matches memories containing "whisper API" and "audio-to-text" even without exact keyword overlap.

Data Storage

~/.lightning-memory/

memories.db # SQLite database keys/ private.key # Nostr private key (chmod 600) public.key # Nostr public key (your agent identity)

Roadmap

• Phase 1: MCP server with local SQLite storage
• Phase 2: Lightning intelligence (vendor reputation, spending summary, anomaly detection)
• Phase 3: Nostr relay sync (NIP-78, Schnorr signing, bidirectional sync)
• Phase 4: L402 payment gateway (macaroons, Phoenixd, HTTP gateway)
• Phase 5: Compliance & trust (budget enforcement, vendor KYC, community reputation, pre-flight gate)
• Phase 6: Memory marketplace (gateway discovery, remote L402 queries, gateway client)
• Phase 7: Agent reliability (semantic search, deduplication, contradiction detection, circuit breakers)

Star History

License

MIT