Why AI assistants lose context mid-conversation and how memory systems can be rebuilt

A developer spent five days stuck in a loop where an AI assistant repeatedly returned the same incorrect configuration for an Airflow DAG after Microsoft moved the service from Azure to Fabric, with little documentation available. The experience prompted a ground-up investigation into why AI assistants fail to retain information across and within conversations. The author identified two distinct problems: a knowledge gap, where the information simply did not yet exist, and a memory gap, where the assistant could not track corrections or prior context. Common fixes like expanding context windows or summarising conversations both proved inadequate — larger windows are costly and models still miss information in the middle, while summaries discard the specific details needed for accurate recall. Research cited from the MemGPT paper found GPT-4 answered questions about past conversations correctly only 32 percent of the time when working from summaries, underscoring that the core issue lies in memory architecture rather than model capability.