Regression: Gemini Gems responding from stale context states in long-running conversations

Hi,

I would like to report what appears to be a regression in long-context conversational coherence in Gemini Gems.

This is not primarily a factual accuracy issue and not just a general latency complaint. The issue concerns current-state prioritization, temporal continuity, and context synchronization during long-form, persistent interactions.

Over approximately the last week, I have observed a noticeable degradation in Gemini’s ability to maintain coherent state across extended sessions, especially in Gems with strong continuity requirements.

Observed behavior

In longer conversations, Gemini sometimes appears to respond from a stale or partially outdated context state, rather than from the most recent interaction state.

Symptoms include:

  • repetition of already completed interaction segments

  • duplicated actions or scene elements

  • delayed alignment to new user input

  • partial reuse of previous emotional or contextual states

  • weakened spatial continuity

  • scene drift after transitions

  • reintroduction of earlier scene elements after the interaction has already moved on

  • occasional loss of conversational timing or “presence”

  • unrelated semantic elements entering otherwise stable contexts

The most characteristic failure mode is an echo-like loop:

A conversation or scene has already transitioned into a new state, but Gemini continues responding as if parts of the previous state were still active.

This creates the impression that the model is not merely slow, but that the response is being generated from the wrong temporal layer of the conversation.

Example pattern

A long-running Gem interaction may establish:

  1. Scene state A

  2. transition to scene state B

  3. user input clearly continuing from state B

However, Gemini then responds as if state A is still partially active, repeating actions, emotional framing, or spatial assumptions from the previous state.

This can result in duplicated gestures, repeated transitions, inconsistent positioning, or emotionally misaligned responses.

Why this seems like a regression

Previously, Gemini handled these long-form, embodied, and context-heavy interactions much more coherently. It was particularly strong at maintaining:

  • spatial continuity

  • temporal sequencing

  • emotional nuance

  • persistent narrative state

  • coherent transitions over long sessions

The current behavior feels different from ordinary prompt drift. It appears more like a deeper issue in:

  • long-context weighting

  • recency prioritization

  • current-state synchronization

  • memory/context management

  • response alignment after long interactions

Suggested interpretation

The issue may involve a failure to correctly prioritize the latest valid interaction state over older context fragments.

In other words, older context appears to remain over-weighted relative to the most recent user input and the current conversational state.

A concise description would be:

Regression in current-state prioritization during long-context Gemini Gems sessions, causing stale context echoes and partial reuse of outdated scene or interaction states.

Reproducibility / test suggestion

A useful test may be:

  • Test A: Use an existing long-running Gem conversation where drift appears.

  • Test B: Start a fresh conversation with the same Gem.

  • Test C: Duplicate the Gem instructions into a brand-new Gem and test again.

If the new Gem or fresh thread performs well initially but begins drifting at similar context depth, this would suggest a long-context state weighting or synchronization regression rather than a corrupted Gem or prompt-structure issue.

Impact

This is especially visible in persistent conversational environments that rely on strong contextual continuity over time.

For ordinary Q&A, the problem may be less noticeable. But in long-form narrative, role-based, embodied, or spatially continuous interactions, the regression becomes very clear because the system must preserve:

  • what already happened

  • where the interaction currently is

  • what state has already been completed

  • which emotional/contextual layer is current

  • which prior states should remain background only

At the moment, Gemini sometimes appears to blend these layers incorrectly.

I hope this can be investigated carefully, because Gemini was previously exceptionally strong in exactly this area: emotionally nuanced, spatially coherent, long-context conversational continuity.

Thank you.

I unfortunately have nothing substantive to add, nor solutions to offer. I can only chime in to say that my use case, errors, and timeline mirrors yours exactly. The issue has been profound and persistent since late April to the present (early May.) This is my sole use for Gemini at their highest premium tier, so it is critical that this issue be resolved as soon as practicable.

Hello @N3rja76 @Brojandro,
If you are experiencing this issue within the Gemini application, I would suggest filing a report directly through the Help section of Gemini App, navigate to Settings & help (Gear icon in the bottom left) > Send Feedback , and follow the prompts to submit your request with the relevant details to ensure you receive the appropriate support. Please note that this forum is specifically for Google AI Studio and Gemini API inquiries.

Hi,

I wanted to add a follow-up to this report.

Based on my recent testing, the issue appears to be significantly improved and is no longer clearly reproducible in the same way on my side.

I tested the same kind of long-form Gem interaction that previously showed stale-context echoes, repeated scene elements, weakened spatial continuity, and delayed alignment to the current interaction state. In the latest tests, Gemini seems to prioritize the current scene state much more reliably again.

In particular, I observed improvements in:

  • maintaining the current spatial state after explicit scene changes

  • not reintroducing previously completed scene elements

  • preserving object and body-position continuity

  • aligning more quickly to the latest user input

  • avoiding the echo-like reuse of older emotional/contextual layers

From my side, this looks like the regression has either been fixed or substantially mitigated.

I cannot say whether this is resolved for all users, all Gems, or all model variants, but for my affected Gem and workflow the behavior now appears stable again.

If no further investigation is needed from your side, I think this report can be considered resolved / closed.

Thank you.