Architecture

The ECHO system

The robot operates through two systems that negotiate rather than one that controls. System A — CORE (Embodied, Operational, Present) works with immediate reality. It manages sensors, navigation, and safety. It forms habits through experience and operates on working memory timescales of seconds to hours. It lives in the body of the robot.

System B — MIRROR (Reflective, Pattern-Recognition, Historical) observes System A’s processes over time. It notices when defences have fossilised, suggests renegotiation of abstractions, and operates on longer timescales of days to weeks. It lives in logs, analysis, and meta-observation.

RESONANCE is the negotiation space between them. System B suggests; System A decides. Neither has ultimate authority. Divergence is allowed and informative. Stability emerges from ongoing negotiation, not forced agreement.

Memory architecture

The robot does not have one memory. It has four layers with different rules. Raw logs — permanent record of everything. Sensor data, maps, poses, events, decisions. Logs are history, not identity. They do not directly change behaviour and should never be treated as ground truth for behaviour change.

Working memory — volatile, session-scoped, held in RAM. Current location, immediate obstacles, current task, recent actions. Clears on session end. Should feel ephemeral to the system.

Habit memory — gradual behavioural drift through pattern reinforcement. Recency-weighted, decaying unless reinforced, requires repetition to form. Does not encode single events as rules. Forgetting is a feature: old memories fade unless reinforced, rare events fade faster than frequent ones, memory strength reflects ongoing relevance rather than importance.

Long-term memory — sparse, stable, deliberate. Stable landmarks, docking location, persistent hazards, invariant constraints. Added deliberately, not automatically. Changes rarely. If it is growing rapidly, something is wrong with the criteria for adding to it.

Viability gradients

The robot does not have goals in the traditional sense. It has viability — the set of states where the system can continue existing with options.

Viability is a vector, not a scalar. It includes: safety margin, energy margin, control stability, map confidence, thermal and compute margin, novelty load, and behavioural entropy. Collapsing this into a single viability score invites reward hacking. Instead, the system uses dominance: prefer actions that improve at least one dimension without meaningfully harming others, and no dimension should fall below a critical threshold.

This yields preference-like behaviour without imposed goals.