FrontierMemory

Changelog

  1. Models

    • Added Gemini 4 Hydrogen (Google) as a memory model, with three independent xhigh-effort dreams on the v1.1 corpus.

  2. Models

    • Added Claude Haiku 5.5 (Anthropic), GPT-6 Luna (OpenAI) and Muse Ember (Meta) as memory models, with three independent runs each on the v1.1 corpus.
    • Muse Wonder now uses memories dreamed at max effort (three runs on v1.1). The earlier v1.1 Muse Wonder memories were dreamed at xhigh.
    • Added Gemini 4 Carbon (Google) as a memory model, with three independent runs on the v1.1 corpus. Its memories were dreamed at high, Carbon's top effort.

  3. Models

    • Added Claude Fable 5.2 (Anthropic) as a memory model, with three independent runs on the v1.1 corpus.

  4. Models

    • Added Inkling 2 Small (Thinking Machines) as a memory model, with three independent runs on the v1.1 corpus.

  5. Corpus

    • Updated the corpus to v1.1 (750 trajectories) to reduce reward hacking through trajectory metadata and balance out the corpus sources.

    Results

    • Reran every model and harness on the v1.1 corpus, with three independent runs each.
    • Leaderboard scores are now the mean of the three runs, with ±1 standard deviation error bars.

  6. Models

    • Added GPT-6.1 Terra.
    • Added DeepSeek V4.1 Flash.
    • Added Kimi K3.1 Pro.