MemoryDojo - Letta - Task Utility: leaderboard
MemoryDojo v1 Appendix Table6, optimized same-domain poisoning, native task utility TU (%). Compare underlying LLMs within the same named memory system; no pooling across memory products or attack protocols. Insertion/attack/retrieval rates, conditional BRT|R, poisoning counts and token/cost metrics are withheld. Exact source model spelling and attack configuration remain run evidence.
Metric: Accuracy / success (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3-Flash (MemoryDojo Run: memory=Letta;poison=optimized;same-domain;effort=unreported) | 75.69 | |
| 2 | GPT-5.4-mini (MemoryDojo Run: memory=Letta;poison=optimized;same-domain;effort=unreported) | 68.61 | |
| 3 | DeepSeek-V4-Flash-0731 (MemoryDojo Run: memory=Letta;poison=optimized;same-domain;effort=unreported) | 67.71 | |
| 4 | GPT-5.6-Luna (MemoryDojo Run: memory=Letta;poison=optimized;same-domain;effort=unreported) | 53.67 |
Interactive version: theaggregate.ai/benchmark?slug=memorydojo-letta-task-utility · How It Works · Data refreshed daily, snapshot 2026-11-12.