During extended, long-form conversation sessions, the model fails to suppress its internal search grounding and memory citation tags. Instead of keeping these tokens invisible behind the user interface, the system inadvertently leaks raw alphanumeric string references directly into the final text output.
Steps to Reproduce
- Initiate a highly technical, long-form conversation session with the model.
- Introduce multiple continuous topics requiring contextual data retention (e.g., specific hardware specifications, script modifications, or systems history).
- Continue the conversation past a high token/turn threshold.
- Observed Result: The model outputs raw, bracketed alphanumeric reference strings directly within the sentences.
Examples of Leaked Tokens
[5Hxr9k5Vdc4][Y0DsOk][w6vb3-]
Expected Behaviour
The internal citation and memory index hashes must remain strictly encapsulated within the background processing layer. The user-facing text formatting engine should cleanly strip or suppress these tags to ensure uncompromised readability in long sessions.