Self-Improvement Memory Enables Continual Recursive Self-Improvement
Xin Yu, Fei Xia, Jiaqi Pan, Youbao Tang, Chengzhi Liu, Adams Yu, Sicong Jiang, Yanfei Chen, Lingzhou Xue, Ting Yu · Nov 2026 · Google Cloud AI, Google DeepMind, Penn State
arXiv 2611.18856
Interactive page
The harness-optimizer loop
Weights stay frozen. Toggle Self-Improvement Memory (SIM) to see what the optimizer can read.
Four conditions, same policy, tasks, rounds, budget
Hover a row. Harness Transfer is the direct control for memory.
Hover a condition.
Two kinds of change in CRSI
Both require a compatible executable interface (old harness still runs).
One round with SIM, step by step
Memory is for the optimizer, not the task agent. Illustrative entry schema
Why log the prediction first?
Hover each card.
Hover a card.
Reported results
Outside the coding paired study, every cell is a single optimization run.
Reported margin ranges
Full SIM minus Harness Transfer, and Problem-Only minus Harness Transfer.
Coding study: 10 paired source runs (Terminus-2, TB2.1 → TB4)
Each source run has two continuations, with and without SIM. The pair is the unit.
Completion-gate case (one illustrative case, not a measurement)
Noise yardstick
The only variance yardstick: spread ≈ 1.8 points between paired coding differences, inferred from the confidence interval, not stated in the paper.
Open question: structure or just more context?
Controls the discussion proposes. Proposed, not run in the paper
Caveat ledger