AI-IQ adds beliefs, dreams, graphs

This is a more ambitious take on agent memory than the usual “embed and retrieve” pattern. If it works as advertised, the real value is not search alone but stateful memory that can reason, decay, and self-correct over time.

• –FTS5 plus sqlite-vec gives it a practical hybrid retrieval stack for offline use without cloud dependencies.
• –Beliefs with confidence scores and prediction resolution are the strongest idea here because they turn memory into something that can be updated, not just appended.
• –Dream consolidation and retrieval feedback tuning are useful for long-running coding agents, where duplicates, stale facts, and noisy context accumulate fast.
• –The tradeoff is complexity: once memory becomes a mini cognitive system, schema discipline and tests matter more than the SQLite packaging.
• –The repo claims production use and hundreds of tests, which is the kind of signal this category needs before it is worth trusting in a real workflow.