When users interact with an LLM, their data flows in two ways: through the context window (temporary session memory, cleared after the conversation) and, if used for training, embedded permanently into the model's weights.
The forgetting capability varies drastically by data type. Session and chat history can be easily deleted. RAG-ingested documents can be removed from vector databases. Fine-tuning data is difficult to erase, often requiring retraining. But core training data is nearly impossible to remove — it's woven into billions of model parameters.
| Data Type | Forgettable? | Method | Key Challenge |
| Session/Chat History | Yes | Deleted from logs/session storage | Minimal - straightforward deletion |
| RAG / Ingested Documents | Yes | Removed from vector database | Requires re-indexing; cached copies may persist |
| Fine-tuning Data | Difficult | Requires retraining or selective unlearning | Costly, time-consuming, may degrade model performance |
| Core Training Data | Near Impossible | Machine Unlearning (experimental) | Data influence embedded in billions of weights; not separable |
Machine Unlearning, an emerging research field, attempts to selectively strip specific data's influence from trained models, but remains immature for large-scale production use.
Regulations like GDPR and India's DPDP Act 2023 grant users a "Right to Erasure," yet technical erasure from model weights remains an unresolved compliance gap for AI companies.
In practice, companies delete chat logs, purge retrieval databases, and exclude data from future training — but cannot fully erase existing model weights.
Moving forward, LLMs can forget stored data, but erasing learned influence from model weights remains one of AI's biggest unsolved challenges.
