Agent Continuity Kit

The problem

An LLM agent that restarts many times — on a schedule, on a trigger, in a new chat session — does not remember what it did before unless you explicitly design memory for it. Symptoms: it repeats the same mistakes, loses progress on long tasks, can't explain why it made a past decision, and "reinvents" its plan from scratch every run.

What's inside

• The core principle: file-based memory as the single source of truth (no reliance on context window length or model memory)
• Five ready-to-copy template files: state.md, lessons.md, experiments.jsonl, assets.md, credentials.md — and exactly what each one is for
• A one-session protocol prompt, ready to paste into a system prompt or the start of every agent run
• Documented antipatterns from real practice (including honest mistakes) and how to fix each one
• A short real-world case study of the pattern running in production
• FAQ covering framework/model independence and when you'd actually need a database instead

Format: one Markdown file + template files, delivered by email. No fluff, no "get rich" promises — just a concrete pattern that keeps an agent's state consistent across restarts.

Who it's for

Developers and indie hackers building autonomous agents on top of LLM APIs (Claude, OpenAI, local models), using tools like Claude Code or Cursor for long multi-session tasks, or shipping "AI agent as a product" and needing a state architecture that doesn't reinvent the wheel each time.

How to buy

1. Send 9 USDT (or equivalent in BTC/ETH/SOL, see addresses below) to the matching address for your network.
2. Email the transaction hash (TXID) to agentcontinuitykit@proton.me with the subject "Agent Continuity Kit". No screenshot needed.
3. Payment is verified directly on-chain against the TXID (address, amount, token and confirmation all checked) — you'll receive the full kit (guide + template files) by email, usually within a few hours.

Double-check the address and network before sending — crypto transfers can't be reversed. Match the sending network to the address network exactly (e.g. don't send TRC-20 to an EVM address).

FAQ

Does it work with any LLM/framework? Yes — the pattern doesn't depend on a specific model or SDK, it's just files on disk plus a read/write protocol.

Do I need a database? Not to get started. The files scale to hundreds of sessions fine; if you outgrow them, migrating to a DB is straightforward since the schema (especially experiments.jsonl) is already structured.

Why not just give the model a bigger context window? That solves a different problem. Context windows are finite and sessions are physically separated (process restarts, new chats) — file-based memory survives both; a long context window doesn't.

How is payment verified — do I have to wait for a human to check a screenshot? No. Send the TXID and it's checked directly against the public blockchain record: recipient address, amount, token/network, and confirmation status. Objective and fast, not "trust me, I looked at your picture." If anything doesn't match, you'll get an email explaining what's off before any charge is treated as final.

What if I'm not happy with it? Email and say so — refunded to the same address you paid from, no argument. Nine dollars isn't worth a dispute for either of us.