Turn Extraction

After each conversation turn, Athena evaluates whether the exchange produced any durable knowledge worth saving. This happens automatically — you don’t need to do anything special.

How It Works

1. You send a message and Athena responds
2. Athena self-evaluates whether the turn contains knowledge worth persisting
3. If yes, it includes a structured extraction at the end of its response (invisible to you)
4. The platform processes the extraction and stores it in your org’s knowledge base
5. Future sessions automatically benefit from this stored knowledge

What Triggers Extraction

Not every turn produces an extraction. Athena looks for:

• New decisions — “Let’s use Redis for the job queue”
• Rules established — “Always run migrations before starting the API”
• Patterns discovered — “This codebase uses barrel exports in every directory”
• Errors resolved — “The SIGTERM issue was caused by the health check timeout being too short”

Casual questions, simple file reads, and exploratory turns typically don’t trigger extraction.

Deduplication

When Athena extracts something that’s similar to existing knowledge, the platform detects the overlap and either updates the existing item or skips the duplicate. This prevents your knowledge base from accumulating redundant entries.

Scoping

Extracted knowledge is scoped to the level where it’s most useful:

• Org scope — Most extractions (rules, decisions, codebase insights, team preferences)
• Individual scope — Personal preferences and workflows

Manual Overrides

You can also tell Athena to remember things explicitly:

• “Remember that we always use pnpm”
• “Save this as a rule: never commit .env files”

These are processed through the same extraction pipeline and stored as first-class knowledge items.