Cortex-engine adds persistent memory for agents

Hot take: this is more compelling than a standard vector-store wrapper because it treats memory as something that can evolve, not just accumulate.

• –Typed memory and belief revision directly address the “append-only bloat” problem that makes many agent memory systems degrade over time.
• –The NREM/REM consolidation framing is a strong product story, but it will live or die on whether it actually improves retrieval quality and reduces noise in practice.
• –FSRS is a smart choice for long-term recall, especially if the implementation is transparent enough for users to trust and tune.
• –The biggest risk is complexity: graph retrieval, consolidation, scheduling, and multi-provider support can become hard to reason about unless the defaults stay genuinely simple.
• –The local-first SQLite + Ollama path is a real selling point for privacy, reproducibility, and low-friction adoption.