Summarize too aggressively and the agent acts on wrong assumptions
If summarization drops a detail that turns out to matter, the agent doesn't fail loudly, it fails silently. It keeps reasoning and acting as if the missing detail never existed, often producing output that looks confident and coherent while being built on an incomplete picture. A user who told the agent 'don't touch the production database' early in a session, only for that constraint to get compressed out of a later summary, may find the agent doing exactly the thing it was told not to do, with no warning that anything went wrong.
This is what makes memory bugs so dangerous compared to other kinds of agent failures. A tool call that errors out is visible, the agent gets an error message and can react. A forgotten constraint produces no error at all, it just produces a wrong action that looks perfectly normal from the outside.