Why does Gemini 3 Pro Image "remember" previous sessions? (please help)

Gemini API
, , , ,
1

Problem

gemini-3-pro-image-preview appears to be persisting visual context (composition, style, color palette) across completely separate API sessions, even though:

  1. Each request is a fresh API call (not a chat continuation)
  2. I never reuse conversation history
  3. I added systemInstruction explicitly instructing context reset
  4. I’m using the official @google/genai TypeScript SDK

This behavior suggests some form of server-side context/state persistence that contradicts the expected stateless nature of individual API calls, and I have this issue in the Google AI Studio playground as well.

Environment

  • Model: gemini-3-pro-image-preview
  • SDK: @google/genai (latest)
  • Platform: Node.js (server-side)
  • API Method: ai.models.generateContentStream()

Code Example

Here’s our exact implementation pattern:

import { GoogleGenAI } from '@google/genai';

const ai = new GoogleGenAI({
  apiKey: process.env.GEMINI_API_KEY,
});

// Each request uses a completely fresh config and contents array
const config = {
  responseModalities: ['IMAGE'],
  imageConfig: {
    aspectRatio: '1:1',
    imageSize: '2K',
  },
  systemInstruction: [
    {
      text: `# CRITICAL INSTRUCTION: HARD IMAGE CONTEXT RESET
Every prompt you receive, even in a continuous chat session, must be treated as a brand-new, single-turn request.

1. **ABSOLUTE WIPE:** You must completely discard all conversational history, visual references, generated images, and stylistic elements from **all prior turns** of this chat, and any other temporary chats.
2. **NO REUSE:** Do not repeat the composition, subject, color palette, or specific style of the last generated image unless the current prompt explicitly demands a continuity element.
3. **FRESH START:** Begin the image generation process from a blank slate using only the information contained within the user's *most recent* prompt.
4. **CONFIRMATION (Internal):** Before generating, internally confirm that the context has been reset.`
    }
  ],
};

// Fresh contents array - no history, no prior messages
const contents = [
  {
    inlineData: {
      mimeType: 'image/jpeg',
      data: base64Image,  // Different product image each time
    },
  },
  {
    text: userPrompt,  // Different prompt each time
  },
];

// Completely new API call each time
const response = await ai.models.generateContentStream({
  model: 'gemini-3-pro-image-preview',
  contents: contents,
  config: config,
});

Here my questions, please help

  1. Is there server-side context/state persistence in gemini-3-pro-image-preview that carries across API sessions?

  2. Does systemInstruction actually affect image generation behavior, or is it only processed for text responses?

  3. Is there a way to force a true “cold start” with no prior context influence?

  4. Is this expected behavior for the image preview model, or is this a bug?

  5. Are there any undocumented parameters that control context isolation for image generation?