DocsTechnology
Memory & Brain
Memory stores what the engine learned about how an app works. Brain is the app’s view of that knowledge; the selected model uses relevant context to plan the next task.
Inspect app knowledge
Settings → Computer → Brain lists the applications the Brain knows. Each one opens its own page, with Graph and List views of the controls, groups and relationships gathered while workers used it. The Brain belongs to the computer: every worker and the command line share it, and this view only reads it. Before any worker has used an app, it shows “No App Knowledge Yet”.
Brain is the knowledge view, not another model provider. The selected model reasons about a task using the context supplied to it.
From interaction to knowledge
When a worker observes a window, the elements of the scene are added to that app’s Brain, scoped to the window’s title.
When it acts, the effect it verified is recorded against the element it acted on. Controls, groups and transitions form app knowledge.
Before acting, the engine asks the Brain what a verb on an element is expected to do, from trusted transitions. The target is still found in the current window.
Learning an interaction is not retraining the model. Stored knowledge describes the app and observed behavior; it does not guarantee that an old route still works.
Use knowledge with live state
- Start with the task and current app context.
- Recall relevant controls, relationships or previously observed effects.
- Let the model choose a next step.
- Find the target in the current window and check the result after acting.
App updates, new dialogs and layout changes can require another observation or revised knowledge. Memory does not authorize an action or make old coordinates valid.
Knowledge, history, session
| Data | Lifetime and purpose |
|---|---|
| App knowledge | One JSON file per application in Knowledge, under ~/Library/Application Support/Mecum, with daily backups kept for 14 days. Inspected through Brain. |
| Workspace history | Workers, conversations, messages and tool events in Workspace.store, in the same folder. |
| Window context | Temporary references and geometry for the current observation and session. Invalid after a restart. |
| Provider session | The provider’s session ID is saved with the conversation and offered again only to the same provider. Start fresh context clears it. Resuming a session depends on whether the provider still holds it. A failed resume is reported rather than replaced with a new session. |
See Workspace for what reopening a conversation restores.
What writes to the Brain
The Brain is written by worker and command-line sessions: what they observe in an adopted window and what their actions changed. Nothing in this release learns from your own activity: the Brain receives only what a session observed and did.
An external MCP client uses the shared Brain and living memory only when granted that capability; otherwise it has its own Brain. The Passive Watcher capability records interaction events only while a client runs it, and those events are not learned automatically.
There is no separate learning switch. A worker with desktop access learns from the apps it uses; revoking its macOS permissions stops new observations without erasing what was saved.
See Permissions & data for the macOS grants and where the files live.
Checked against Mecum app source at main 524eb7f on October 1, 2026. UI labels and paths have not yet been checked against the signed release build.