Learning Without Remembering: Habituation in Weights and a Self-Model That Acts in a Minimal Embodied Agent

We present a minimal embodied agent that runs without vision or audition, on body channels only: an action-conditioned body predictor, a homeostatic drive (E,C,U,S), episodic memory, and a meta-cognitive module Φ that predicts its own prediction error. A pre-registered 30-seed study tested whether this loop can detect violations of its sensorimotor contingencies, habituate by updating its model, and retain that update in weights rather than in explicit memory. Detection was reliable (z=20.6, 95% CI [16.0, 25.5]). Habituation cut surprise by 86% (d=3.5). The trace lived in the weight delta (post/pre ratio 0.02). The meta-cognitive module Φ was calibrated (r=0.701), generalized out of distribution (r_cross=0.730), and causally changed behavior: the agent left an unreliable zone faster when Φ was coupled to action (15.0% vs 28.1% time, d=-1.61). One pre-registered control failed (C3) and is reported in full. All code, registration notes, and raw logs accompany this preprint; the complete system runs on a MacBook Pro (M4 Pro, Apple silicon) with no discrete GPU. This is version v0.12. For reproduction see README.md and requirements.txt in the files. Companion code includes LFM2.5-1.2B used as a frozen translator under the LFM Open License v1.0.