Mistral Vibe: Smarter Compaction That Preserves Your Conversation

How Mistral Vibe Handles Long Conversations

Mistral Vibe is designed for extended coding sessions: multi-step refactors, exploratory debugging runs, and back-and-forth problem solving that can span dozens of turns. As sessions grow longer, the underlying model context fills up and Mistral Vibe triggers compaction, summarizing conversation history into a compact representation so the session can continue without hitting token limits.

Prior to v2.13.0, compaction worked by replacing the active conversation with the generated summary. The existing message history was effectively discarded and the summary stood in its place. This approach had a subtle but impactful flaw: if compaction ran again in the same session, the summary from the first compaction could be misidentified as a user message. Subsequent summaries would then include confusing artifacts like "Last request from user was: [previous summary text]", a corruption that compounded with each additional compaction cycle.

What Changed in v2.13.0

Mistral Vibe now injects the compaction summary into the conversation rather than replacing it. With injection, the conversation structure remains intact and the summary is woven in as context rather than substituted wholesale. Repeated compaction cycles no longer risk cascading corruption, because the mechanism correctly distinguishes between genuine user messages and injected summaries.

Why This Matters for Long Sessions

Developers who regularly use Mistral Vibe for extended sessions often encounter compaction multiple times in a single sitting. The previous replacement behavior meant that very long sessions could degrade in quality over time, as the model's view of the conversation became increasingly muddled by misattributed summary content. The injection approach keeps the conversation grounded, so compaction operates as a background memory management mechanism rather than a disruptive reset.

Additional Changes in v2.13.0

Mistral Vibe v2.13.0 also fixes a crash that could occur during initialization, and moves the VS Code extension promo banner to the top of the conversation view rather than pinning it above the input field. The extension, now published as mistralai.mistral-vibe-code, is accessible via updated links in the CLI's onboarding and plan-upgrade prompts.

Key Takeaways

• Compaction summaries are now injected into the active conversation rather than replacing it, preserving the conversational thread.
• Repeated compaction no longer corrupts message history, fixing a bug where prior summaries were treated as user messages.
• Long-running coding sessions become more reliable, as the model sees a coherent conversation thread even after multiple compaction events.
• The fix addresses a compounding failure mode where each additional compaction worsened session memory quality.
• No user-facing configuration change is required; the improvement applies automatically to all sessions.
• The improvement complements the custom compaction prompts feature shipped in v2.11.1.