Hello Google Cloud Community and Gemini Team,
I would like to propose an essential feature for the Gemini web interface and API ecosystem that addresses the issue of automatic context eviction in long sessions.
The Problem:
Currently, when the context window limit is reached, the system automatically evicts the oldest messages. This often leads to the loss of critical initial instructions, core data, or source code discussed at the very beginning of the chat session, rendering the remaining conversation less accurate.
Proposed Solution (User-Controlled Memory Cleansing):
Introduce a feature that allows users to manually select which parts of the chat history to delete from the active memory pool:
- Visual Anchors: Automatically add a small index number (1), (2), (3), etc., next to every user and AI message in the chat history.
- Context Alert: When the context window utilization reaches 90%, display a system notification: “Memory is almost full. Please select which messages to evict to preserve core context.”
- Manual Eviction: Allow users to click a “trash/forget” icon next to specific index numbers, or type a direct command in the chat box (e.g., “/forget 3 5 9 12”).
- Under the Hood: The platform completely wipes the text of the selected messages from the active token attention window, instantly lowering token usage while keeping vital anchor messages (like system prompts or initial code blocks) permanently intact.
This feature would give developers, prompt engineers, and heavy users absolute control over long-form AI interactions. I would love to hear the community’s thoughts on this!
Best regards,
Velax31