generateTitle silently fails with reasoning models (gpt-5-mini, o1, o3) due to hardcoded temperature: 0

[lzj960515]

Summary

generateTitle silently falls back to "Conversation" when the agent model (or the generateTitle.model override) is a reasoning model such as gpt-5-mini, o1-mini, or o3-mini. The failure is swallowed by a try/catch with only a DEBUG-level log, making it very hard to diagnose.

Root Cause

createConversationTitleGenerator in agent.ts calls generateText with temperature: 0 unconditionally:

const result = await context.traceContext.withSpan(llmSpan, () =>
  generateText({
    model: resolvedModel,
    messages,
    temperature: 0, // reasoning models reject this parameter
    maxOutputTokens,
  }),
);

Reasoning models (OpenAI o-series, gpt-5-mini, etc.) do not accept the temperature parameter. The AI SDK emits a warning and the call either errors or returns no usable output. The surrounding try/catch catches this silently:

} catch (error) {
  context.logger.debug("[Memory] Failed to generate conversation title", {
    error: safeStringify(error),
  });
}
return fallbackTitle; // always "Conversation"

Steps to Reproduce

1. Create an agent with a reasoning model, e.g. model: "openai/gpt-5-mini"
2. Configure generateTitle: true on the Memory instance
3. Start a new conversation
4. Observe: conversation title is "Conversation" instead of an AI-generated title
5. In logs: AI SDK Warning: The feature "temperature" is not supported for reasoning models

Expected Behavior

• generateTitle should detect reasoning models and omit unsupported parameters (temperature)
• Or the error should be surfaced at warn level so users know why title generation failed
• The docs should note that reasoning models are not supported for generateTitle

Workaround

Explicitly override the model in generateTitle config to use a non-reasoning model:

new Memory({
  storage: adapter,
  generateTitle: {
    enabled: true,
    model: "openai/gpt-4o-mini", // must be a non-reasoning model
    systemPrompt: "Generate a short title (max 6 words).",
    maxLength: 60,
    maxOutputTokens: 24,
  },
});

Environment

• @voltagent/core latest
• OpenAI provider, model gpt-5-mini (reasoning model)

[omeraplak]

Thanks again for the clear report and repro details. I opened a fix in #1249.

The fix keeps the existing default behavior (temperature: 0) for compatibility, adds generateTitle.temperature: null so reasoning models can omit the parameter, and surfaces unsupported temperature/title-generation failures at warn level with guidance.

We will get this reviewed and released soon.