Memory, work profile, and global rules

Put context in the right place

Write the current deliverable in the conversation. Keep lasting facts and preferences in Settings → Memory. Put recurring instructions in Settings → Instructions & rules. A model connection does not, by itself, configure memory.

See Design systems for reusable visual guidance.

Memory injection and learning are separate

1. Open Settings → Memory and inspect Enable memory injection in the header.
2. Enable it to make saved context available to later tasks. Turning it off stops injection and the associated automatic flows.
3. Check Learn from chats separately. It captures useful facts from conversations; it is not a promise to retain every chat verbatim.

Turning injection off does not delete existing entries. Pause injection for a temporary break; edit an entry to correct a particular fact.

Create and maintain an entry

1. Choose Add or import memories, then Add manually → New memory.
2. Set Title, Type, Description, and Content. Use the description as a short summary and include the scope in the body.
3. Choose Create, or Save when editing. This form is not auto-saved; Cancel or closing it does not submit a draft.
4. Return to Saved memory. Filter by type, open Preview, and use Edit to correct outdated information. Confirm what you need to retain before deleting.

Title: Audience and writing defaults
Type: User
Description: Delivery conventions for product documentation
Content: Our readers are product managers and designers.
Use English for documentation; retain exact UI button labels.
Apply to help articles and product explanations.

Work profile and connected-app imports

In Add or import memories → Work profile, describe your role, team, domain, audience, taste, deliverables, language, and current goals. Choose Save profile. These fields supply task context; they are not an invitation form.

Import from apps lists connection options. Connect the required account, inspect the proposed facts, and save only relevant information. Authorization, scanning, and saving are distinct steps. Resolve a missing connection or access restriction in the source app. Never put passwords or API keys in a memory entry.

Automatic hooks and the memory model

How it works explains the flow. Depending on the version, separate hooks control profile context, request enrichment, result verification, and learning from chats. The master switch sets the overall boundary; each hook has its own state.

Memory model under Models & providers selects a model for memory processing when required. It is separate from the task model in the composer. If memory processing fails, check that model’s authentication and allowance even when ordinary chat still works.

Use Advanced to inspect saved records, extraction history, and the detailed memory structure. Before reimporting, check the master switch, learning state, saved entry, and memory-model configuration.

Set recurring instructions

1. Open Settings → Instructions & rules.
2. Write concrete, compatible requirements in Global rules. Keep project-specific constraints in the project.
3. Reopen the section to confirm the text, then check the behavior on the next task. Existing work is not rewritten retroactively.

Preserve approved copy and brand assets when revising existing work.
Explain changed files and validation when delivering code.
State assumptions when essential business information is missing.

When memory seems missing or too broad