Meta FAIR Self-Improving Memory Scores 56.0 on Continual Learning Bench vs 44.6 for Best Hand-Written Harness
Meta FAIR's Self-Improving Memory lets Kimi-K2.6 write its own LOAD and OFFLOAD memory prompts, then distills them into its weights. Optimized prompts score 56.0 on Continual Learning Bench, versus 44.6 for the best hand-written harness, according to MarkTechPost.
Meta FAIR's Self-Improving Memory method scores 56.0 on Continual Learning Bench with optimized prompts, versus 44.6 for the best hand-written harness, according to a MarkTechPost report published November 19, 2026.
How the method works
The approach has Kimi-K2.6 write its own LOAD and OFFLOAD memory prompts. These prompts are then distilled into the model's weights.
Reported results
| Setup | Continual Learning Bench score |
|---|---|
| Optimized prompts (Self-Improving Memory) | 56.0 |
| Best hand-written harness | 44.6 |
The figures are as reported by MarkTechPost. Details such as the benchmark's task composition, the number of evaluation runs, and any other baselines are not covered in the summary available for this article. Parameter count, context window, and pricing for the setup are not stated.
What this means
The reported comparison is between memory prompts written by the model itself and the best hand-written harness, both measured on Continual Learning Bench. On that benchmark, the model-written prompts score higher. The distillation step, as described, moves the model-authored LOAD and OFFLOAD prompts into the model's weights rather than leaving them only in the prompt layer. Teams building memory harnesses by hand have a reported reference point of 44.6 against which the 56.0 figure was measured.
Related Articles
Anthropic: Selective notes beat full logs for Claude agent memory, with 'Dreaming' fixing errors
Anthropic published research on how Claude retains knowledge across sessions using note files. Its tests show selective notes outperform exhaustive archives, and a post-session process called Dreaming can correct faulty notes. All results come from Anthropic's own tests.
Ai2 launches Theorizer in Asta and open-sources the AI agent that proposes scientific theories
Ai2 launched Theorizer, an AI agent that combines findings from scientific papers to propose explanations for observations, inside its Asta platform. The code is open-source, and Ai2 covers token costs up to a limit in the web interface. A lung cancer case study shows the workflow.
Salesforce's BLIP-Live video agent cuts unnecessary responses by over 50%, company claims
Salesforce AI Research introduced BLIP-Live, a general model for live video agents that decides when to speak and when to stay silent. According to Salesforce, in one evaluation across 326 ego-centric videos it reduced unnecessary responses of state-of-the-art models by over 50%.
Tunix agentic GRPO lifts Gemma 4 E2B from 9.4% to 95% win rate in Plane Strike
Google's developer blog details how Tunix's agentic GRPO pipeline trained Gemma 4 E2B to play Plane Strike with tool calls. The win rate on held-out boards rose from 9.4% to 95% after 100 steps on a single Cloud TPU v6e-8 host.
Comments
Loading...