Why Your AI Has the Memory of a Goldfish (and How to Fix It)

Large language models (LLMs) often forget instructions over the course of a conversation, similar to how people with ADHD struggle to maintain focus. This is due to the attention mechanism in transformer architecture, not a bug. The solution is not bigger context windows or better prompts, but building in structured checkpoints to validate outputs against the original requirements.

Details

Large language models (LLMs) like GPT and Claude can initially follow detailed instructions, but then start drifting away from them after 8-9 turns. This is not a bug, but a result of how the attention mechanism in transformer architecture works. Early tokens carry more weight, but as the conversation grows, those initial instructions get pushed further from the model's active attention. This is similar to how people with ADHD can forget instructions that are not immediately reinforced. Trying to fix this by repeatedly pasting the instructions is a coping mechanism, not a real solution, as it burns up context window space. The fundamental issue is architectural - longer context windows don't solve the underlying attention weighting problem. The true fix is building in structured checkpoints to validate the model's outputs against the original requirements, similar to how external systems and routines help people with ADHD manage their attention deficits.