Understanding the Active Token Limit for Gemini’s “Saved Info” Feature

As users of advanced AI tools like Gemini 2.5 Pro, we often uncover nuances that official documentation may overlook. One such discovery involves the “Saved Info” feature, which is integral for tailoring our interactions with the AI. Recently, I’ve encountered an issue that seems to reveal an interesting limitation—one that I suspect could benefit from community insight.

Key Insights from My Experience

During my extensive use of the “Saved Info” feature, I’ve populated it with numerous instructions, including my persona, formatting preferences, and personal details. However, I’ve identified a few critical observations that shed light on how the feature operates behind the scenes:

1. Invisible Limitations: It appears that the system has an implicit limit on the number of active instructions. When I add more instructions than this threshold, the AI starts to “forget” the oldest entries. Although these instructions are still visible on the settings page, they seem to be excluded from the active context, creating a silent truncation effect.

2. Increased Response Time: As the number of active instructions grows, I’ve noticed a corresponding increase in latency. This delay in processing each prompt seems reasonable, given the heightened computational demands placed on the system.

My Current Understanding

Through my experimentation, I liken the relationship between “Saved Info” and context utilization to the dynamics of a computer’s storage systems:

• “Saved Info” as Hard Drive Storage: This feature appears to have vast, near-infinite capacity, storing extensive data in the background.
• Context Window as RAM: The active chat operates within a significant context, estimated at around 1 million tokens.
• The Critical Bottleneck: There appears to be a limited “startup file” of Saved Info that is actively loaded into memory for each session. It is this particular file size that I seek to quantify.

The Quest for Answers

This leads me to a pressing question: Has any member of the community pinpointed a specific number regarding this active limit? Are we discussing a token threshold of around 4k, 16k, or even 32k? I’m eager to hear if anyone has quantified the maximum size of that “startup file” based on their experiences.

I was contemplating conducting a “canary test”—adding instructions incrementally until I identify the limit—yet I thought it prudent to first gather insights from fellow users who