AI Agent Forgetting Context? You're Not Passing History

This one isn't a model limitation — it's almost always a code problem. Every LLM API call is stateless by default. If your agent only sends the latest message instead of the full running conversation, of course it "forgets" — it never received the earlier messages in the first place.

This is easy to miss early on, because the first couple of exchanges work fine (there's not much history to forget yet), and the bug only becomes obvious a few messages in.

The fix:

every API call needs to include the full message array — every prior user and assistant turn, not just the newest one. Most frameworks give you a conversation/messages array specifically for this; the fix is almost always "actually pass it," not a deeper architecture change.