Tejas Parthasarathi Sudarshan has a point. Machines don’t forget. They can’t. And that limitation runs deeper than any glitch in code or oversight in design. It sits at the core of how neural networks store information across weights, context windows, retrieval systems and persistent logs.
Sudarshan laid this out plainly in his July 10, 2026, essay on tejassuds.com. The piece dissects four distinct memory systems in modern AI. None of them decay like human recollection. Weights embed training data so thoroughly that extracting specific influences remains an open problem. Context windows flip between total recall and total erasure. Retrieval stores pull old embeddings with the same sharpness as fresh ones. Logs and backups turn intentional erasure into a complex distributed systems challenge.
Short. Brutal. Accurate.
But the implications stretch further. They touch privacy laws, model safety, corporate liability and the very nature of human-AI interaction. Recent research and regulatory pushes show the issue gaining urgency. Companies and governments want AI systems that comply with the “right to be forgotten.” Yet the technical path forward stays messy. A new paper accepted to ICLR 2026 examines “Machine Unlearning under Retain–Forget Entanglement,” highlighting how efforts to remove data often tangle with what the model must keep. LinkedIn post by Ping LIU announced the acceptance in early 2026.
And then there’s the human side. Forgetting serves a purpose. It enables generalization. It drains emotional charge from painful events while preserving facts. Machines lack any equivalent mechanism. Their memory is read-only. Retrieval pulls vectors untouched. No reconsolidation. No adaptive pruning.
Sudarshan draws on classic studies to make the contrast clear. Hermann Ebbinghaus tracked forgetting curves in the 1880s. Richards and Frankland argued in their 2017 paper that transience helps brains avoid overfitting to details. Reconsolidation research from Nader, Schafe and LeDoux in 2000 showed how recall itself modifies memories in humans. None of these processes exist in current architectures. Tejas Sudarshan’s essay ties them together with striking figures showing human decay curves against machine cliffs of perfect retention or catastrophic loss.
Catastrophic forgetting itself reveals the flip side. When models learn new tasks they often overwrite old knowledge wholesale. McCloskey and Cohen named the phenomenon in 1989. Transformers also exhibit “Lost in the Middle” behavior, per Liu et al. in 2023, where information in the center of long contexts vanishes from attention. These aren’t features. They expose architectural gaps.
But the forgiveness angle cuts sharper. Sudarshan defines it in engineering terms. Keep the record at full fidelity. Stop letting it drive behavior. Leave the option open to act on it anyway. No current method achieves that. Machine unlearning tries to erase influence, closer to amnesia. Time-to-live rules evict on schedules. Fine-tuning applies blanket changes. None offer per-memory choice with full awareness.
Recent pushes in research and regulation reveal how companies and labs wrestle with these constraints.
IBM Research published an explainer on October 6, 2024, titled “Machine unlearning for LLMs.” It describes methods that target weights to influence long-term memory without full retraining. The post notes that true erasure guarantees remain elusive. IBM Research blog.
Google’s December 2025 work on Titans and MIRAS takes a different tack. It treats certain forgetting mechanisms as regularization to balance new learning against past knowledge. The architecture aims for test-time memorization with surprise metrics, letting models update core memory on the fly. Google Research blog.
Even newer material from 2026 shows momentum. A February 5 piece in Communications of the ACM discussed machine unlearning as a way to remove sensitive information like voices or copyrighted material more efficiently than retraining. It referenced techniques for AI-generated voices developed by researchers at Sungkyunkwan University. CACM article.
The International Association of Privacy Professionals ran an analysis on February 25, 2026. Nicoletta Kolpakov wrote that generative AI makes memory probabilistic across billions of parameters. Removing one person’s data requires altering the model’s fundamental patterns. She calls machine unlearning the technical frontier for reconciling human rights with these systems. IAPP analysis.
Legal and ethical pressure keeps building. The European Union’s General Data Protection Regulation and similar rules elsewhere assume deletion is straightforward. AI makes it anything but. A 2025 arXiv survey on machine unlearning classifies approaches into traditional methods, verification techniques and domain-specific strategies. It runs over 100 pages and catalogs hundreds of papers. Progress exists. Yet verification that unlearning actually worked without side effects stays hard. arXiv survey.
One 2026 arXiv paper proposes “unlearning by design.” Models get trained with key deletion mechanisms from the start. The MUNKEY approach decouples instance-specific memorization from core weights in memory-augmented transformers. Early results suggest better control. But it requires redesigning systems rather than patching existing ones. arXiv paper on key deletion.
Critics push back too. A NeurIPS 2025 paper later expanded in October 2025 argues machine unlearning doesn’t do what people think. Authors including A. Feder Cooper and Kate Lee point out mismatches between goals like copyright removal or safety and what implementations actually deliver. They warn against overpromising to policymakers. Paper PDF.
So what now?
Engineers face hard trade-offs. Storage costs keep dropping. That removes the biological pressure that forced human brains to prune. Without artificial scarcity or new computational primitives, graceful decay won’t emerge naturally. Some labs experiment with synthetic budgets or regularization that mimics metabolic expense. Others focus on retrieval systems that attach timestamps or decay factors to embeddings. Results remain preliminary.
Business leaders should watch closely. AI deployed in customer service, hiring, lending or therapy carries perfect records of every interaction. A model that never softens its view of past complaints or errors could erode trust. Companies might face lawsuits not just for what they remember but for how that memory shapes outputs years later.
Researchers at AllenAI used their open OLMO model and Dolma corpus in late 2025 to test unlearning in practice. They found not all data unlearns equally. Some facts resist removal more than others. The work underscores why open models matter for studying these limits. AllenAI blog post.
Yet the core insight from Sudarshan’s piece holds. We built systems with perfect fidelity because it was easier. Cheaper. Faster to scale. Now we live with the consequences. Second chances in human societies grew from our hardware limitations. Biology gave us mercy by default. Machines received no such gift.
Fixing it demands more than bigger context windows or better alignment techniques. It may require rethinking memory itself. Introduce cost. Build reconsolidation analogs. Create architectures where recall can modify stored representations safely and selectively. Until then, we deploy entities that have never experienced the first mercy. Forgetting.
And that changes everything about how we interact with them.
WebProNews is an iEntry Publication