mlsanityData to 14 Oct 2026

AI Engram: In Search of Memory Traces in Artificial Intelligence

arXiv 2606.14997 ↗First posted 12 Jun 2026cs.AI cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Memory formation is fundamental to intelligence, yet whether deep neural networks preserve identifiable memory traces analogous to biological memory units remains an open question. This work introduces a geometric framework to identify such "AI engrams" by formalizing the neuroscientific criteria of specificity, reactivation, sufficiency, and necessity into a constrained inverse problem. We derive a closed-form estimator that isolates individual memory traces from globally entangled parameters, and show that this biologically-derived solution corresponds to a natural gradient update on the parameter manifold. AI engrams enable surgical manipulation of learned knowledge: any subset of memories can be composed or erased through linear arithmetic, without iterative optimization. Experiments ranging from simple MLPs to LLMs demonstrate the causal validity and substantial scalability of AI engrams. Together, these results bridge theories of biological memory and artificial representation learning and offer geometric insight into how deep networks simultaneously support functional specificity within distributed storage.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 41 of its 54 references that are in the map. First posted on arXiv 12 Jun 2026.

Builds on

41 of 54 references.

References · 41

  1. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  3. 2024The Llama 3 Herd of Modelscited by 14,715
  4. 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,375
  5. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  6. 2019Similarity of Neural Network Representations Revisitedcited by 1,755
  7. 2018World Modelscited by 1,513
  8. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  9. 2022Editing Models with Task Arithmeticcited by 1,120
  10. 2019Machine Unlearningcited by 1,061
  11. 2022Mass-Editing Memory in a Transformercited by 1,001
  12. 2022Toy Models of Superpositioncited by 971
29 more
  1. 2017PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruningcited by 898
  2. 2015Optimizing Neural Networks with Kronecker-factored Approximate Curvaturecited by 796
  3. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 708
  4. 2019Linear Mode Connectivity and the Lottery Ticket Hypothesiscited by 637
  5. 2019Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networkscited by 618
  6. 2024TOFU: A Task of Fictitious Unlearning for LLMscited by 533
  7. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 509
  8. 2024Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearningcited by 504
  9. 2023Unified Concept Editing in Diffusion Modelscited by 388
  10. 2023SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generationcited by 386
  11. 2017Wasserstein Auto-Encoderscited by 373
  12. 2021Unrolling SGD: Understanding Factors Influencing Machine Unlearningcited by 279
  13. 2023Towards Unbounded Machine Unlearningcited by 268
  14. 2024Rethinking Machine Unlearning for Large Language Modelscited by 257
  15. 2019What's Hidden in a Randomly Weighted Neural Network?cited by 237
  16. 2021Machine Unlearning of Features and Labelscited by 229
  17. 2023Model Sparsity Can Simplify Machine Unlearningcited by 214
  18. 2025Open Problems in Mechanistic Interpretabilitycited by 179
  19. 2024Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearningcited by 172
  20. 2023Editing Implicit Assumptions in Text-to-Image Diffusion Modelscited by 130
  21. 2023Task-Specific Skill Localization in Fine-tuned Language Modelscited by 105
  22. 2024What makes unlearning hard and what to do about itcited by 89
  23. 2022Reliability of CKA as a Similarity Measure in Deep Learningcited by 87
  24. 2025OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metricscited by 83
  25. 2024UNDIAL: Self-Distillation with Adjusted Logits for Robust Unlearning in Large Language Modelscited by 50
  26. 2024Alternate Preference Optimization for Unlearning Factual Knowledge in Large Language Modelscited by 33
  27. 2025Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decompositioncited by 27
  28. 2025Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocolscited by 18
  29. 2025Stochastic Parameter Decompositioncited by 16

Built on by

Not cited yet.