> ## Documentation Index
> Fetch the complete documentation index at: https://agent-harness101.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Why not one blob

> Each format is chosen for how that state gets written, read, and trusted.

## The question

Why not just keep one JSON blob per session instead of seven stores?

## The answer

Each format is chosen for how that state gets written, read, and trusted.

### JSONL — conversation & subagent history

Append-only, so a crash never corrupts prior turns, and it's replayable.

### JSON — task state

Current status, not history, so it's overwritten and queried compactly.

### Markdown — persistent memory

Durable lessons need to stay readable and editable by a human.

### Separate subagent transcripts

Isolation at the storage layer, so exploration noise can't leak into the parent.

### Code stays in the repo

One source of truth; a copied snapshot would go stale the moment the file changes.

The filesystem itself is the lightweight persistence layer.

## Continue the walkthrough

<Card title="Token levers" icon="arrow-right" href="/cost/token-levers">
  Continue to topic 6 of 14.
</Card>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.