EditorBrain can remember what you like and what you want fixed, so the next agent has that context. A correction does not change your preferences and does not become a permanent rule by itself.
From your comment to a decision
- You say something specific: "The title covers my face; leave space above the presenter."
- The agent saves the observation as project feedback and can capture a suggested memory for that project.
- It continues the correction according to your review mode, without inventing an approval.
- When you review the suggestions, you decide whether they apply only to the video, whether they should be kept for the brand, or whether they should be discarded.
In Memory you can review the proposals and their scope. Approving the video or liking a resource does not confirm a permanent memory. "This turned out well" also does not authorize changing the settings of all your clients.
The API does not automatically read your conversation with Claude or Codex. The agent has to send the feedback; the skills and tools tell it to do that. If the comment stayed only in the chat, another agent cannot recover it from EditorBrain. You can ask: "Check that you saved my corrections as suggestions for this project."
Scope matters
| Scope | When to use it |
|---|---|
| Project | A decision that belongs only to that video or job |
| Style | An adaptation note for that style, also tied to a brand |
| Brand | A recurring preference you confirmed for that brand |
Conditions are kept too: "in this series," "when I show an app" or "only in ads" does not mean "in every video." Memories are not shared between clients. Saving a suggestion does not by itself invalidate the editorial context of a render; confirming an effective rule can require refreshing it.
When it contradicts a setting
For example, "no word-by-word captions" can clash with a saved karaoke treatment. EditorBrain flags conflicts it recognizes and shows the proposal; it cannot automatically interpret every free-form sentence. The agent should explain the difference and ask about scope when that is needed.
Accepting the memory does not change the structured setting. If you also want to change the caption treatment, the agent needs that instruction and saves the preference through its usual operation. The same split applies to music, effects, format and render or storage.
Save from the CLI or MCP
Available in CLI and skills 0.51.0 and the compatible API:
editorbrain feedback add --project my-project --text "The title covers my face" --capture-memory --category layout
editorbrain memory list --scope project --id my-project
In MCP, use save_editorial_feedback with projectId, text, captureMemory:true and the category that applies. Capture returns a note and the suggestion status; do not announce that a memory was created if the response indicates a limit or a warning. Repeating the same observation keeps its identity and does not reactivate a discarded suggestion.
When you submit a corrected version:
editorbrain production submit --project my-project --file renders/revision-v2.mp4 --title "Title moved" --parent latest --addresses "The title covers my face"
--parent latest requires a single candidate review; if the variants are ambiguous, use its UUID. An alternative, a technical change or a continuation uses the matching revisionKind, without inventing user feedback. MCP shares this contract in submit_production_review.
After you save or confirm remote memory, the CLI follows the returned localSync and runs pull --project my-project, keeping conflicts. An MCP-only agent reads the context again. It does not overwrite local documents to hide differences.
Check what is pending and what was learned
The full context returns the text, conditions, scope and origin of pending memories. The first response can include up to 12: pendingCount, pendingTruncated and pendingScopes say whether you need to fetch the rest. The orientation summary does not contain every entry.
Improvement analysis needs real feedback and must be enabled by the owner. Two or more recent notes can give material for suggestions; they do not guarantee a useful rule and they do not turn on automatic training. The user keeps the final decision, and consent to improve the product is independent.
See also preferences, Studio and project continuity.
See how memory was used in a video
In Memory, choose a project under “Memory in your projects”. You can see the confirmed memories included in its context and the decisions the agent recorded: applied, adapted, omitted, or exception. Each decision keeps its version and reason. A rule can be adapted to the video without changing the saved preference.
“In context” does not prove application. “Declared by the agent” is not visual verification either. When a review is linked, its availability and approval are shown separately; a link does not prove that every memory was followed. Historical decisions remain even after the context changes. An empty history means no decisions have been recorded yet, not necessarily that the agent ignored your preferences.
Your agent can use list_memory_usage over MCP or
editorbrain memory usage --project ID with the CLI. Both read the same history.