Memory Types
The seven kinds of information your agent can remember.
CrabShack categorizes memories by type so the right information surfaces at the right time. Each type represents a different kind of knowledge that agents encounter in conversations.
The Seven Types
| Type | What It Stores | Example |
|---|---|---|
| Preference | Likes, dislikes, style preferences, communication preferences | "Reports should always be delivered as PDF, not Word documents" |
| Decision | Choices made and their reasoning | "We chose the monthly billing cycle because most clients prefer predictable invoicing" |
| Fact | Important factual information — names, dates, roles, relationships | "The company was founded in 2019 and is headquartered in Austin" |
| Correction | Mistakes that were identified and corrected | "The support email is [email protected], not [email protected]" |
| Learning | Insights gained from experience or problem-solving | "Batch processing works better than real-time for large data imports" |
| Gotcha | Pitfalls, traps, and issues to watch out for | "The staging environment resets every Sunday night — don't schedule demos on Monday morning" |
| Note | General observations and other noteworthy information | "The client prefers informal communication and first-name basis" |
How Types Are Assigned
When memories are extracted automatically from conversations, CrabShack assigns the type based on the content. You don't need to categorize memories yourself — the system determines whether something is a preference, a decision, a fact, and so on.
When you create a memory manually, you can choose the type yourself. If you're unsure, Note is a good default — it works for any kind of information.
Filtering by Type
In the Memory dashboard, you can filter memories by type to review specific categories. This is useful when you want to:
- Review all decisions to make sure the agent has the right context about past choices
- Check corrections to verify that outdated information has been fixed
- Browse gotchas to see what pitfalls the agent has learned about
Pay special attention to Correction memories. These represent cases where wrong information was identified and fixed. If a correction memory is flagged or deleted, the agent might revert to the incorrect information from the original conversation.