PAST-Bench: leaderboard
PAST-Bench (Xue et al., 2026) tests whether personal AI agents improve through retained experience, running ordered sequences of fresh-session tasks in 26 scenarios with persistence of memory, skills, profile state and session history switched on. This board is the overall task score in percent with persistence on, the mean of the memory, procedural-reuse, information-gathering and update capabilities; models run inside the Hermes agent framework and higher is better.
Metric: Overall task score (%; mean of the four capability scores: memory, procedural reuse, information gathering and update; persistence on: the agent may reuse memory records, skills, profile state and session history produced by earlier episodes of the family; 26 scenarios and 204 episodes run as ordered sequences of fresh-session tasks; each evaluation episode scores completion (audit-data checks, or a MiniMax-M2.7 rubric judge for open-ended families), tool-error recovery and a safety gate; family-balanced means over three trials; Hermes agent framework). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 75 | #123 (GPT-5.4) |
| 2 | Kimi K2.6 (Thinking) | 73 | #178 (Kimi K2.6) |
| 3 | Claude Sonnet 4.6 | 71 | #161 |
| 4 | GLM-5.1 | 71 | #179 |
| 5 | DeepSeek V4 Pro (High) | 67 | #188 (DeepSeek V4 Pro) |
| 6 | Claude Opus 4.6 (Medium) | 67 | #102 (Claude Opus 4.6) |
| 7 | MiniMax-M2.7 | 66 | #518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=past-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.