Agent Usage

This guide helps agents decide what to store and where.

Before storing

Run through three checks before you store anything:

1. Will this matter tomorrow? If not, skip it.
2. Do you already know this? Search first with recall.
3. Can you cite a source? If yes, use learn. If not, use remember.

Which tool

You have	Tool
Observation, no source	remember
Claim with source	learn
Conclusion from facts	believe
Current reasoning	reason
Understanding changed	reflect

Decision tree

flowchart LR
    A{"Worth storing?"} -->|No| B["Skip"]
    A -->|Yes| C{"Have source?"}
    C -->|Yes| D["learn"]
    C -->|No| E{"Conclusion from facts?"}
    E -->|Yes| F["believe"]
    E -->|No| G["remember"]
    
    style A fill:#d29922,stroke:#d29922,color:#1f2937
    style B fill:#1f2937,stroke:#f78166,color:#f0f6fc
    style C fill:#58a6ff,stroke:#58a6ff,color:#1f2937
    style D fill:#1f2937,stroke:#3fb950,color:#f0f6fc
    style E fill:#58a6ff,stroke:#58a6ff,color:#1f2937
    style F fill:#1f2937,stroke:#3fb950,color:#f0f6fc
    style G fill:#1f2937,stroke:#3fb950,color:#f0f6fc

Common mistakes

No evidence on learn. Calling learn("API is fast") without a source is wrong. If you don’t have a source, use remember instead.

Belief without about. Calling believe("Use caching") with no about is a guess, not a belief. You need to link to the facts that support it.

Expecting reason to persist. The Intelligence layer is session-scoped. If you need the conclusion later, write it to Memory or Knowledge.

Deleting instead of superseding. Pass supersedes when updating to preserve the history chain.

Silent belief changes. When you update a belief or correct a mistake, call reflect so the change is recorded.