A developer is arguing that AI memory systems are fundamentally broken because they treat memory like a search problem rather than mimicking how human brains actually work. In a telling example, his AI assistant persistently recommended Bun.js as a runtime solution for six months based on a casual January conversation he never followed through on. The memory had an importance score of 8/10, but the system had no way to know he'd abandoned the idea two days after mentioning it.

The core issue: append-only memory systems store everything with equal freshness regardless of age. A decision made in week one sits at the same priority level in week eight, even if it was reversed weeks ago. Meanwhile, contradicting memories filed later never accumulate enough access frequency to override the outdated data. The AI confidently retrieves information the user long ago invalidated, and it can take multiple attempts to notice the pattern.

The proposed solution involves lifecycle fields in the database schema: decay_score (0-1, starting at 1.0), confidence (reliability rating), contradicted_by (links to superseding memories), and expires_at (time-based expiration). The system uses exponential decay with a 30-day half-life — memories not accessed for months fade toward zero, and anything below 0.1 gets archived rather than deleted. Frequently accessed memories earn a freshness bonus capped at 1.0. The implementation runs on plain SQLite with no embedding models or third-party APIs required.