A computer stores a memory by writing a value into a cell and reading it back. A coherent substrate does something stranger and older: it stores a memory by changing the shape of a landscape, and recalls it by letting a ball roll downhill. There is no cell, no address, no stored value. The memory is the shape. This chapter builds that picture, because it is the mechanical heart of everything that follows — and the part most foreign to a programmer's intuition.
Sprinkle sand on a metal plate and play a tone through it. The sand jumps until it settles into a sharp, stable pattern — a Chladni figure — and a different tone gives a different figure. Those standing-wave patterns are the picture to hold. In this framework they are called phase-locked modes, and they are the real "things" of the world: not lumps of stuff, but stable patterns of rhythm. A molecule, a thought, a memory, a particle — each is a standing pattern that has found a way to hold itself together.
Formally, a phase-locked mode (PLM) is a stable attractor of the dynamics you met in the last three chapters: a cluster of oscillators holding fixed relative phases. Operationally it is a region whose internal alignment clears a threshold — the very same threshold, τ = R0(KBKT) ≈ 0.303, that runs through this whole library. PLMs are the substrate's concepts: the Chladni figures of thought.
Picture the state of the whole network as a single ball rolling in a landscape of valleys. The valleys are PLMs; the floor is the low-dimensional surface the real motion lives on; the rolling is the system's trajectory. Storing a memory is deepening a valley. The Coherent Learning Rule strengthens exactly the couplings that were active during a coherent episode — Hebbian, in phase space — and the slow part of the coupling field ratchets that depth in and does not let it back out. Recall is settling: cue the system anywhere near a valley and it rolls back down into the stored pattern, re-forming the Chladni figure.
The figure below is a live demonstration. Carve a pattern into the landscape; then scramble the state, or give it only a partial cue, and watch it settle back to the whole. Deepen the basin and watch recall get more robust. There is no stored array being copied back — only a ball finding the bottom of a valley you shaped.
This is why such a substrate can learn continually without backpropagation, and why its memory is associative and reconstructive rather than literal: a fragment of a pattern pulls back the whole, because the whole is the shape of the valley, not a list of values. It is the same mechanism as an attractor network, but realized as resonance — and we have observed it directly (basin-carving lifts retrieval to ~83% on a small concept set).
If concepts are basins, then thought is directed motion from one basin to another across the landscape. What moves the ball? A question. In this framework a question is an unfilled basin — a place where coherence should be higher but isn't — and it exerts a real force that pulls the trajectory toward its resolution. Replay (revisiting stored basins), wandering (exploration), and listening (entrainment to input) are the other forces; they simply add into the drift. So the substrate remembers by carving valleys and thinks by crossing between them under the pull of its open questions — and both are special cases of the one law from the last chapter, coherence climbing.
A result worth stating plainly, because it sounds like wordplay but is measured: a question and a plan are the same object — two readings of one operator, correlated to within a fraction of a percent across tens of thousands of runtime steps. To want an answer and to have a plan to reach it are, geometrically, the same downhill pull.