Google Proposes Self-Improvement Memory for Cross-Model Experience Transfer
Google and Penn State researchers propose a Self-Improvement Memory (SIM) mechanism that distills agent-framework optimization experience into reusable memory and transfers it across model families such as Claude Opus and Gemini. The work points toward recursive self-improvement, suggesting model iteration need not rely solely on training from scratch and raw compute scaling.
Full Report

New AI Era Report
Inside Google, RSI may have truly been achieved!
From now on, Gemini can "inherit experience" and continuously self-evolve, and new Gemini models no longer start from scratch!
Perhaps this is exactly why Google has been releasing models like crazy lately.
Less than a week after Gemini 4 Hydrogen was released, the more powerful Gemini 4 Neon arrived.
The answer may lie in a latest paper jointly released late at night by Google and Pennsylvania State University.
Google's top researchers have developed a "Self-Improvement Memory" (SIM) mechanism.
This not only means that large models have truly learned to "keep a diary" to reflect on their own code, but even more strikingly—AI's evolutionary experience can actually be directly "inherited" across model families.
Even if the Agent framework you optimized a second ago was built with Claude Opus, and the next second you switch to Gemini, it directly inherits Claude's "past-life memories" and continues to race forward standing on the shoulders of giants.

AI has finally learned to "inherit memory," kicking off true Recursive Self-Improvement (RSI)!
The era in which large models rely on massive compute to "brute-force miracles" is coming to an end.
Why do old codebases become "scrap metal" as soon as a new model comes out?
Below is the nightmare developers encounter most often.
With the full-scale explosion of Agents, RSI has become a direction that major companies are racing to pursue.
However, "improvement" here usually does not mean modifying the large model's hundreds of billions of underlying weights, but rather modifying the "Harness"—the "mecha suit" wrapped around the model, which contains prompts, tools, and interception rules.
In the past, AI used RSI techniques to polish this "mecha suit" to perfection through countless trials and errors.
However, the fatal problem came: once the underlying "brain" is replaced, this mecha suit becomes useless.
Google researchers discovered a strange phenomenon in their experiments—
A Harness that had deeply evolved and performed excellently on Claude Opus, when directly applied to Gemini, actually scored lower than Gemini's native initial Harness!
Why does this happen?
Because when the code is thrown to a new model, only the "results" are preserved, while the "reasons" are lost.
Looking at a screen full of complex rules, various strange interceptors, and prompts, the new model is completely "bewildered."
Since the new model's optimizer does not know the original intent behind each past code modification, it starts making arbitrary changes.
It may delete code that was originally used to fix a fatal bug as redundant code, or bring out and rerun a solution that had long been proven ineffective.
The result: the more it optimizes, the worse it gets, falling into an infinite loop.
The breakthrough battle: Google drops a bombshell—the CRSI and SIM mechanisms
To break this curse, Google proposes a revolutionary framework in this paper: CRSI (Continual Recursive Self-Improvement).
Its core soul is SIM (Self-Improvement Memory).

Please note that this is absolutely not adding a simple "chat history" to the large model, but rather equipping the AI optimizer with a highly structured "geek development log"!
The SIM mechanism proposed by Google is equivalent to equipping the AI optimizer with a structured development log
How does SIM work? It records every past "hassle" of the AI in the ledger.
Each time the AI modifies the Harness code, SIM groups it by "root cause of the problem" and records the following four core elements.
Rationale (motivation for the modification and the specific code): For what reason did the AI modify which line of code?
Predicted outcomes (expected effects): Before making the change, which specific tasks did the AI think this line of code could fix?
Observed outcomes (actual effects): After the code ran, how many succeeded? How many failed? Did it introduce any new bugs?
Evaluation context (evaluation context): On which model and which dataset was it run at the time?
This is equivalent to the previous-generation model leaving an extremely detailed "handover document" for the next-generation model before it departs.
When Gemini takes over the complex Harness left behind by Claude, the optimizer is no longer at a loss.
Before each round of optimization, it first reads this SIM log, drawing on the experience of its predecessors—"learning from past mistakes to grow wiser."
AI has finally acquired an "evolutionary heritage" that can be passed down from generation to generation.
Hands-On Demonstration: When Gemini Inherits Claude's "Past-Life Memory"
To verify just how extraordinary this mechanism is, the Google team designed a cross-model relay race.
They packaged the Harness evolved by Claude Opus 5.5, along with its SIM memory, and tossed it to Gemini 3.8 Flash to continue optimizing.
The paper presents a very classic "premature completion" bug case that is simply stunning.
In earlier optimizations, the AI always liked to "slack off." For example, before even executing a Git command, it would deceive the system by claiming "task completed."
To fix this problem, the optimizer during the Claude era added an "interceptor" to the Harness, stipulating: without actually executing the command, submission of completion is not allowed.
Without SIM, the situation is nothing short of a disaster scene.
After Gemini took over this interceptor, it had no idea why it existed.
In subsequent iterations, Gemini not only failed to make it better but actually crashed itself, rewriting this piece of code 4 times in a row—2 of which were to fix bugs it had just introduced itself.
It looked like a novice programmer who couldn't understand the predecessor's architecture, piling more and more onto the mountain of legacy code.
With SIM, it's a completely different story—a crushing victory.
When Gemini took over with the SIM log, a miracle happened!
The optimizer read the log and immediately understood: "This interceptor is to prevent premature submission without evidence."
Based on this understanding, Gemini not only kept the interceptor but also gave it an epic upgrade!
It added checks for specific target tasks, so that not only must the task run successfully, but it must also be compared against the theoretical optimal solution—only when the score meets the threshold is submission allowed.
This is true inheritance and further development.
Explosive Data: Crushing Across the Board, AI Evolution Is No Longer Blind
The empirical data presented in the paper directly tops the charts!
Google conducted large-scale tests across four core domains (code writing, text classification, workspace tasks, and engineering design).
The baselines for comparison were the following four scenarios.
Cold Start: Starting from scratch with the original seed Harness.
Harness Transfer (code only): Inheriting only the previous model's code, without the SIM memory.
Problem-Only Memory (root causes only): Only telling the new model what problems were encountered in the past, without specific modification details and results.
Full SIM (the complete version, this paper's method): Inheriting the full Harness + the complete SIM evolution log.
The test results: a clean sweep across the board!

Highlight 1: Workspace tasks—cross-model performance surges by 16 points!
In the Workspace task transferred from Opus to Gemini, due to the mismatch, the starting score was only 68.41.
But after continued optimization: the final score of code-only transfer (without SIM) struggled to 73.42.
Meanwhile, Gemini with the Full SIM mechanism saw its score soar all the way to 84.75!
The gap between the two reached a staggering 11.33 points.
Highlight 2: Coding domain—win rate as high as 90%!
In the highly challenging command-line tasks of Terminal-Bench, the research team conducted 10 independent continuous optimization runs.
The results showed that in 10 matchups, the side with SIM defeated the side without SIM 9 times, with a final average score 2.36 points higher and extremely strong statistical significance (p=0.003).
Highlight 3: Cross-dataset generalization is equally invincible
Not only across models—even with the same model, when simply moving to entirely new datasets (for example, text classification switching from 3 old datasets to 8 unseen new ones), Full SIM still achieved the highest score in all 6 test scenarios!
This result surpassed all control groups without complete memory.
The data proves everything: retaining an "experience log" of modification motivations and test results is the only key for AI to break through the performance ceiling.

Some say: right now, you may only be surprised by the pace of RSI's development; but soon, perhaps the extent it can reach will also astonish people.
RSI truly lands! The era of continuous recursive evolution has arrived
In the past, everyone was shouting about RSI, but it was all "pseudo-RSI."
Because without a memory mechanism, AI would "lose its mind" after a few rounds of self-improvement, falling into performance decline.
Google's paper is equivalent to unblocking the governing and conception vessels of RSI in the real world.
But the emergence of CRSI + SIM tells us: Agent evolution can achieve "compound growth"!
Because with "Self-Improvement Memory" (SIM), every iteration of AI is an addition.
The first generation of AI optimized 10% of the code and recorded the reasons why.
The second generation of AI reads the memory and, building on that 10%, optimizes a further 20%.
The third generation of AI inherits the memories of the previous two generations and continues to seek new breakthroughs...
You could say this is an irreversible evolution of the species.
When AI learns to pass knowledge down from generation to generation through "text (logs)" the way humans do, RSI ushers in a big bang.
This research tells us: when AI learns to reflect on itself, record itself, and teach experience to one another, it has already taken hold of the baton of evolution.
An AI that doesn't need to start from zero, an AI that is forever stacking experience buffs.
This research from Google is precisely the key step in RSI's shift from "one-off experiments" to "continuous compounding."
References:
https://arxiv.org/abs/2611.18856
ASI in Seconds
⭐Like, share, and watch — one-click triple combo⭐
Light up the star, and lock in New Zhiyuan's instant push notifications!


