Google's Self-Improvement Memory Enables Agents to Evolve Across Models Continuously
Google researchers introduce Self-Improvement Memory (SIM), a mechanism that records optimization rationales and outcomes for agent harnesses, enabling continual recursive self-improvement across model and task changes, with experiments showing SIM outperforms framework-only transfer in cross-dataset and cross-model settings.
Background and Motivation
Google Cloud AI, Google DeepMind, and Pennsylvania State University researchers published the paper Self-Improvement Memory Enables Continual Recursive Self-Improvement (November 13, 2024). The work addresses a practical problem: when an agent's harness (the framework managing context, tools, and execution flow) is optimized for one model (e.g., Claude Opus 5.5), directly transferring that harness to a different model (e.g., Gemini 3.8 Flash) often degrades performance below the new model's original harness. The optimization history — why a rule was added, what failure it fixed, which alternatives were tried — is lost, making further adaptation difficult.
Self-Improvement Memory (SIM) and Continual Recursive Self-Improvement (CRSI)
SIM is a persistent, updatable record attached to the harness. Each entry is organized by recurring root-cause categories and stores:
The code change and its rationale
The task the change aimed to fix
Expected outcome (recorded before evaluation) and actual measured outcome
Domain, benchmark, model, and iteration identifiers
During each optimization round, the proposer and selector components read the full SIM. Proposals must link to an existing root cause or introduce a new one. Only accepted modifications and their measured results are written back; rejected proposals leave SIM unchanged. When the task or model changes, both the harness and its SIM are carried forward, allowing the optimizer to retain, adjust, replace, or delete mechanisms while preserving historical context. This loop is termed Continual Recursive Self-Improvement (CRSI).
Experimental Setup
Experiments covered four domains: programming, text classification, office tasks, and engineering design. Two transfer scenarios were tested:
Cross-dataset (same model): Six setups, each optimized for 10 iterations with 5 candidates per iteration.
Cross-model: Harness first optimized around Claude Opus 5.5, then transferred to Gemini 3.8 Flash for continued optimization on the same benchmarks. Proposer and selector used Gemini 3.8 Flash in both phases.
Four baselines compared: from-scratch optimization, framework-only inheritance, framework + root-cause inheritance, and framework + full SIM. All shared the same starting harness, target model, iteration budget, and candidate count.
Results
Cross-Dataset Transfer
Full SIM achieved the highest final score in all six setups, leading framework-only by 2.43–4.35 points. However, retaining only root causes (without full history) sometimes hurt: in the AHE programming setup, root-cause-only scored 18.79 vs. framework-only's 20.61.
Cross-Model Transfer (Claude → Gemini)
The transferred harness initially underperformed Gemini's original harness in all four domains, confirming that model-specific adaptations do not transfer automatically. After continued optimization:
Office tasks showed the largest gap: framework-only reached 73.42, root-cause-only 79.22, full SIM 84.75 (starting from 68.41).
To test reproducibility, 10 independent source-phase optimizations on Terminus-2 were each split into SIM and no-SIM branches on TB4. SIM won 9 of 10 pairs, average advantage 2.36 points (95% CI 1.05–3.68, paired p = 0.003). One pair favored no-SIM by 0.61 points.
Sequential Model Generations (Gemini 3.5 Flash → 3.8 Flash → 4 Argon)
CRSI led independent optimization by 3.5 points on 3.5 Flash and 1.6 points on 3.8 Flash, but trailed by 0.8 points on 4 Argon, indicating that accumulated experience can sometimes become a liability.
Shared Framework Across Multiple Models
On text classification, a single harness optimized using aggregated feedback from Grok, Opus, and Gemini scored 75.24 on average, versus 75.20 for per-model optimized harnesses and 73.28 for the original shared harness.
Optimizer Model Sensitivity
Swapping the source-phase proposer/selector model from Gemini to Opus improved independent evaluation scores by 1.06 and 0.05 points on two programming harnesses, showing that optimizer capability matters independently of SIM.
Limitations and Open Questions
The authors acknowledge that historical experience is not universally beneficial; cases where inheritance underperformed from-scratch optimization appeared. Identifying outdated experience and controlling memory growth require longer-horizon experiments. The central question remains: as AI participates in its own improvement, can each trial leave more usable experience for the next?
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
