UC Berkeley researchers Immanuel Abdi and Akshat Gupta showed that evolutionary training wipes out what language models already know, calling continuous learning after deployment one of the biggest gaps in AI P². A 16 November arXiv preprint, not peer-reviewed, reports 83.2% retention across five shifting CartPole environments using those same evolutionary methods, but paired with a physics-indexed memory that files each learned policy by the conditions it was trained under S¹.
The preprint calls its system the Evolutionary Neuroplastic Agent, or ENA S¹.
My read: I would not bet on ENA replacing PPO or SAC in any production setting yet. The 83.2% figure is real but narrow, measured on CartPole only, and the authors are candid that ENA adapts more slowly than PPO and returns less than SAC on HalfCheetah. The idea worth watching is the physics-indexed memory: tagging a learned policy by the regime it was trained under, then retrieving it with Bayesian inference when conditions recur. That is a clean architectural idea that could survive even if the gradient-free wrapper does not.
How the memory works
ENA has three moving parts. First, it evolves small neural networks without gradient descent, the optimisation method behind nearly every modern AI model S¹. It mutates and selects networks through evolutionary search, the same family of methods that UC Berkeley's Abdi and Gupta found can cause catastrophic forgetting in language models P².
Second, when ENA learns a policy that works, it archives that policy alongside a label describing the dynamic regime, the physical conditions of the environment at the time S¹. Think of it as filing a solution in a folder named after the problem it solves.
Third, when the environment shifts back to a regime ENA has seen before, it retrieves the matching policy through Bayesian policy reuse, a probabilistic method that weighs which archived solution best fits the current conditions S¹. The agent does not relearn from scratch. It pulls the old file.
What the numbers say, and what they do not
In a 30-replicate ablation on non-stationary CartPole, the complete ENA architecture achieved a mean success rate of 83.2% across five test environments S¹. Among the variants that learned the harder CartPole environments, the full architecture showed the most consistent retention. It held onto old skills better than stripped-down versions that removed parts of the memory system S¹. A selection ablation confirmed that the retrieval rule itself, not merely the presence of a memory, affected performance S¹.
The caveats are equally concrete. In the evaluated sequences, ENA adapted more slowly than PPO on CartPole S¹. PPO is a standard gradient-based reinforcement learning algorithm. On HalfCheetah, a harder locomotion benchmark, ENA achieved lower returns than SAC S¹. SAC is another gradient-based method.
The authors explicitly state that their evidence does not isolate gradient-free optimisation as the cause of retention S¹. They also state it does not establish superiority over other continual-learning methods S¹. The system was tested on only two simulated environments, CartPole and HalfCheetah, with no real-world robot validation claimed S¹.
Where ENA sits in the wider effort
Catastrophic forgetting has attracted a steady stream of open-source work. The GitHub repository which47/LLMCL, created in February 2024 with 36 stars, analyses and reduces the problem in parameter-efficient fine-tuning of language models P³. A separate repo, SirNosh/Neuroplastic-COT, offers a canonical implementation of Elastic Weight Consolidation and Synaptic Intelligence, two established forgetting-prevention techniques, under an MIT licence P⁴. Stanford and UC San Diego researchers built PhysMem, which scales test-time physical memory for robot manipulation and has gathered 266 stars since March 2026 P⁵. Developer interest in agent tooling on GitHub runs broader still, as a free AI agent tutorial that crossed 82,000 GitHub stars showed in October.
The broader agent reliability problem is active territory. As we found when AI agent teams skipped safety checks they pass alone, systems that behave well in controlled tests can break in ways that are hard to predict under real conditions. And as a prompt injection defense that cut agent attack rates from 37.8% to 4.4% showed, the gap between a benchmark number and a deployable system is where most of the work lies.
ENA differs from these efforts by working in the gradient-free evolutionary setting rather than in fine-tuning or test-time adaptation. For a robotics lab weighing whether to try it, the honest test is whether the physics-indexed memory helps in a setting with more than two regime switches. The paper's 83.2% figure comes from five test environments in CartPole; a lab running a manipulator through six or seven grasp conditions would need to check whether Bayesian retrieval still pulls the right policy.
The next checkpoint is peer review, which has not happened S¹. A comparison against established continual-learning baselines like Elastic Weight Consolidation, which the authors did not run, would be the first thing to watch for.
Sources: S1 — Mitigating Catastrophic Forgetting with Physics-Indexed Specialist Mem · P2 — arxiv.org · P3 — which47/LLMCL · P4 — SirNosh/Neuroplastic-COT · P5 — haoyangli16/PhysMem
Related reading
- Prompt injection defense cuts agent attack rate from 37.8% to 4.4% — our technology desk, 2026-11-03
- Free AI agent tutorial crosses 82,000 GitHub stars — our technology desk, 2026-10-10
- AI agent teams skip safety checks they pass alone — our technology desk, 2026-11-16
Written from 5 sourced items, 4 of them primary.