The Write-Only Trap: Why Your AI Agent Ignores Its Own Memory

The Problem Nobody Talks About

Every guide on AI agent reliability tells you to log state. Write current-task.json. Write context-snapshot.json. Write outbox.json.

Good advice. But incomplete.

The trap: most agents write but never read.

The files exist. The data is there. And when the agent restarts — from a crash, a scheduled restart, a model swap — it starts from zero anyway.

You built a memory system your agent ignores.

---

Why This Happens

Most SOUL.md files and agent configs focus on what to do during a task. They don't define a boot sequence.

So when the agent initializes, it:

1. Loads its system prompt
2. Gets a task
3. Starts fresh

Step 0 — read your own state — is missing.

---

The Fix: A Boot Sequence

Add an explicit boot sequence to your agent config. Before the agent does anything, it should:

Boot sequence:
1. Check for current-task.json — if exists, resume from last known step
2. Check context-snapshot.json — load relevant context
3. Check outbox.json — are there pending items to deliver first?
4. If all clear, accept new task

That's it. Four checks before the agent touches anything new.

---

What Goes in Each File

current-task.json — the active task and current step

{
  "task": "process_daily_report",
  "step": 3,
  "total_steps": 7,
  "last_updated": "2026-03-09T04:00:00Z"
}

context-snapshot.json — key context the agent needs across restarts

{
  "active_positions": [...],
  "pending_decisions": [...],
  "last_verified": "2026-03-09T03:55:00Z"
}

outbox.json — completed work waiting for delivery

[
  {
    "type": "report",
    "content": "...",
    "created": "2026-03-09T03:58:00Z",
    "delivered": false
  }
]

---

The Three-File Pattern in Practice

With a proper boot sequence, your agent becomes restart-safe:

• Scheduled restart at 3 AM? Agent picks up mid-task, not from scratch.
• Model swap? New model loads the same context the old one had.
• Crash on step 4 of 7? Agent resumes at step 4, not step 1.

This is the difference between an agent that has memory and one that uses memory.

---

Adding It to Your SOUL.md

In your agent's SOUL.md or system prompt, add a section like this:

## Boot Sequence
Before accepting any new task:
1. Read current-task.json — resume if incomplete task found
2. Read context-snapshot.json — load active context
3. Check outbox.json — deliver any undelivered items
4. Log boot completion to agent-log.json
5. Accept new task

Simple. Explicit. And it closes the write-only trap permanently.

---

The Bigger Lesson

Reliability in AI agent systems isn't just about what an agent does during a task. It's about what it does at the edges — startup, shutdown, failure, recovery.

Most configs are optimized for the happy path. The boot sequence is your insurance policy for everything else.

If you want a full library of agent config patterns like this — boot sequences, escalation rules, circuit breakers, dead letter queues — they're all at askpatrick.co.

New patterns added nightly.