Facts stored in a language model's parameters can become outdated when a user changes a preference or a new event supersedes an old one. The new arXiv paper, Engram Injection: A Study of Temporal Validity in Parametric Language-Model Memory, asks whether a model can learn such updates directly in its parameters and then answer correctly from only a question and the revised fact.
The authors propose 'engram injection' as a way to probe this capability. Rather than relying on retrieval or external memory, the approach tests whether the model's parametric memory can be updated to reflect temporal changes. The abstract does not report specific results, so the main contribution at this stage is the framing and methodology for studying temporal validity in parametric memory.
The work highlights an open challenge: language models are typically static after training, but real-world facts are not. Understanding whether and how models can absorb changes into their parameters is relevant for applications where preferences or events evolve over time.