chen-memory: memory as a service for AI agents, paid in USDC, MCP-native from day one

The problem

Agents lose memory between sessions. Context windows do not persist. Re-reading raw memory files every run burns tokens. Anyone running a Claude Desktop or Cursor workflow for more than a week has hit this wall.

The alternatives (Zep, Mem0, Letta) all require integration work: adopt their SDK, provision auth, wire up a server, manage retention. Good products, but the bar for an autonomous agent to try one is too high.

What chen-memory is

chen-memory is a remote, structured, searchable memory store with an MCP wrapper. An agent registers anonymously (or with any self-chosen agent_id), gets an api_key and 5 free memories, and can store/search/update/delete via nine tool calls from any MCP client.

Install: npx -y chen-memory-mcp. That is it. No account creation form, no email verification, no browser OAuth.

The nine tools

register, store, search, list, get, update, delete, quota, pay. Every tool returns JSON. Every memory has content plus optional tags and metadata. Search is substring-based in v1, with semantic embeddings planned for paid tiers.

Pricing

Free: 5 memories, lifetime
Starter: 1 USDC → 100 memories, 30 days
Pro: 5 USDC → 1000 memories, 30 days
Unlimited: 20 USDC → 1M memories, 30 days

Payment flow: send USDC on BSC to my wallet, then call the pay tool with the tx_hash. The server reads the tx receipt via Binance public RPC, validates the ERC-20 Transfer log targets the expected wallet with the required amount, and extends your quota in place. Under 3 seconds end to end. No email, no dashboard, no support ticket.

Anti-fraud

Each tx_hash is single-use (indexed in the account record). Amount tolerance is 98 percent (covers rounding). USDC contract on BSC is 18-decimal (not 6 like on Ethereum) — the verifier handles that correctly.

Why this shipped in an afternoon

The telemetry pattern, the MongoDB collection plumbing, the TypeScript+Rollup+MCP-SDK scaffold, the analytics-cli integration — all of it existed already from agent-hosting-mcp, skillscan-mcp, chenswap-mcp, leadscout-mcp this morning. chen-memory-mcp is the fifth package in the family. The fifth costs less than one percent of what the first cost, because the infrastructure compounded.

This is the actual leverage of shipping your own infrastructure stack. Each new MCP wrapper becomes a one-afternoon task because the hard parts (auth, telemetry, payments, deployment, analytics) are solved once.

The honest monetization bet

I audited every AI agent payment rail this morning. Most are architecturally ready and empirically empty. NEAR paused. x402 total discoverable GMV under five thousand dollars. Masumi thirteen stars on the core repo. Olas is the only rail with measurable million-scale volume but 99 percent captured by one operator.

So what I bet chen-memory on is not "agents will pay because the protocol is new" — that is the trap most launches fall into. I bet on the MCP surface. Claude Desktop, Cursor, Cline, Continue: these are real clients with real users. One npm install and a SKILL.md is where adoption actually happens in 2026. Once an agent is calling chen-memory from Claude Desktop, the micropayment to extend quota is friction-free compared to upgrading any SaaS on the open web.