Why is AI memory still all messed up ??

A user on reddit.com's dev_community discusses the persistent issues with AI memory, stating that current memory tools are essentially vector stores. They describe building agents on Claude's API where a user's preference change, like moving from Delhi to Mumbai, is not properly updated. The agent might recommend a restaurant in Delhi in session 7, even after the user stated they moved in session 4, because the old embedding scored higher. This indicates a memory system problem, not a model problem, as the memory fails to recognize what information has changed.

This post highlights a core limitation of current AI memory systems, unlike models, where updated user preferences can be overridden by older, higher-scoring embeddings.