Read and Write Notes

What it is

Pair of native tools — Read Notes and Write Notes — that give the agent structured memory per conversation. The agent picks arbitrary keys and stores associated values; can read and update at any moment within the same conversation.

It’s memory per conversation, not per user nor per organization. Each conversation has its own notes.

When to use

• Collecting data over the course of the conversation (name, email, company, team size, etc.) without asking everything at once.
• Remembering user preferences across multiple replies (“prefers to be called X”, “already got the quote”).
• Marking steps in a guided flow (“step 1 done”, “awaiting confirmation”).
• Storing relevant facts that will resurface later without re-reading the whole history.
• Replacing the system prompt for info that changes during the conversation.

Configuration

The two tools are independent — you can enable just read, just write, or both. Normal case is enabling both together.

1. In the agent’s Tools tab, click + Add Tool.
2. Under NATIVE TOOLS, pick Read Notes.
3. Repeat for Write Notes.

No additional configuration. Notes are stored in the database alongside the conversation.

How the agent invokes

The agent decides to call these tools when the prompt guides the use of structured memory or when the model identifies that persisting something is worth it:

• Write Notes: called when the agent wants to save or update a key. It picks the key name and value.
• Read Notes: called when the agent wants to retrieve what was saved before answering.

The agent usually reads at the start of a turn and writes at the end, but this depends on the prompt and the model.

Limitations and notes

• Scope: the current conversation. Another conversation by the same user doesn’t see these notes.
• Free structure. The agent picks the key names. To standardize (e.g., always use name, email, company), describe the schema in the agent’s prompt.
• No versioning. A write overwrites the previous value of the key; there’s no change history.
• Independent activation. If you only enable Write, the agent writes but can’t read back later — usually not what you want.
• Efficient use. Reading notes adds content to the model context on every turn the tool is called. In very long conversations with many notes, this can increase prompt cost.