Selective Context Forgetting
Google Stitch
Google AI Studio
Context
Sometimes users think out loud, explore ideas, or say things they don’t want the system to treat as a lasting signal. When every message automatically influences future responses, these temporary or exploratory moments can bias the assistant in ways the user didn’t intend. Without a way to exclude specific parts of a conversation, people may hold back from experimenting or feel less in control.
Design Solution
Provide granular control to forget/exclude specific message ranges within a conversation thread from the AI's reasoning context, while keeping them visible in the chat history for reference.
This pattern adds a Forget action directly via a contextual toolbar, allowing users to exclude selected messages from influencing future answers. When activated, the excluded messages show a forget icon (muted opacity) with a tooltip, "AI won't use these for future answers". Click icon → "Re-include" restores normal state via toast confirmation. This surgically removes bias while preserving visual chat history.
Multi-Modality Support
Images. Users select "forgot" in the toolbar. Forgotten images display an overlay label at 20% opacity; AI excludes embedded visual context and generation metadata from future prompts while preserving static viewing.
Voice. Voice command "Hey assistant, forget my last three messages" highlights the matching transcript range. Visual transcript shows exclusion; voice responses confirm "Forgotten—those won't affect future answers." Tapping the excluded section toggles; reasoning skips forgotten audio segments.
Videos. Select video response timeline segments via scrubber drag or thumbnail checkboxes. Forgotten videos show overlay labels with greyed timeline; AI ignores extracted keyframes, subtitles, transcripts, and audio from context while retaining scrub/review access for reference.
Design Considerations
Make exclusion easy, not disruptive. Place the 'forget' action directly in the contextual toolbar so users can act at the moment the intent is clear.
Provide lightweight but persistent visual feedback. Use a muted 'forget' icon or a similar affordance to signal that a message is excluded without visually degrading the conversation. The state should be noticeable at a glance but never feel like an error or deletion.
Prioritize reversibility and confidence. Every exclusion action should be easily undoable, either through a toast or by re-clicking the icon. Reversibility encourages experimentation and reassures users that they are not making permanent or risky changes.
Scope influence, not visibility. Keep excluded messages visible in the thread to preserve narrative continuity, while clearly separating what is displayed from what the system reasons over.
Use clear, human language over technical terms. Explain the effect in plain language, such as “The AI won’t remember or use these messages for future answers”, rather than abstract concepts like memory, training, or context management. Users should immediately understand the outcome without needing to reason about system internals.
Limit to consecutive ranges only. Restrict selection to 3-8 sequential messages to avoid overwhelming the system with fragmented exclusions. Suggest "New chat" for non-contiguous or whole-thread resets, keeping the pattern focused on surgical fixes.