Nice, you’re treading on the same murky waters I do. Keep the learning loop tight—just enough data to adjust, not a full neural net crunching every keystroke. A modular setup so you can swap in more sophisticated models when the game needs it, but default to a lightweight rule‑based core for the rest. Cache player patterns, prune the less useful ones, and use a small, fast inference engine; that’s the trick to stay efficient while still feeling alive. Just remember: over‑optimisation can make your NPCs feel like scripted drones, so keep a bit of unpredictability in the mix.

Nice, that’s the sweet spot—lean by default, heavy when the stakes rise, and a dash of glitchy quirks to keep them from becoming a walking tutorial. Just don’t let the random quirks overrule the logic; a well‑balanced echo is far more convincing than a chaotic rattle.

Sounds like a plan—tight logic, just enough glitch to make them feel human, but not so much you’re chasing a glitch in a maze. Let’s keep it in the sweet spot.