Compression & Path-Closure — the Memory-Attractor Case
Recurcline · α-trace · Basin-Relative Observability
When is a memory an attractor?
A memory has two lives. There is the moment you use it — a single cue and the whole pattern reassembles — and there is the rest of the time, when it sits in the wiring doing nothing. This note is about that difference, and about a sharper question underneath it: what must a system be able to do before it can compress many experiences into a few reusable patterns at all?
The precise claim: for recursive compression to occur, a system's closed loops must be able to carry a lasting effect around them — the condition the kernel names T0-χ (Path-Closure Sensitivity). That condition is necessary but not sufficient. We test it in the one domain where the compression is not inferred but directly on display: recursive memory systems (Hopfield-type attractor memory, and biological memory).
Candidate correspondence
V1–V2 · not empirical alignment
Downstream
Evidence class (per § 14): candidate structural correspondence, V1–V2. Kernel anchor: Claude_Kernel_5_21_2026, v2.35.1. Downstream of canon — applies constraints, does not redefine them.
Why this is the privileged case
Across the surveyed domains, most (geometric/Berry phase, protein folding, hysteresis) let you observe only the loop non-neutrality — the T0-χ residue — and compression is inferred indirectly. The memory-attractor case is different: compression is the function of the system, so it appears as a first-class observable — many episodes collapsing into few stored patterns, and the collapse leaving a durable structural imprint the kernel names directly.
The collapse itself is recurcline (RC, § 8.2): the derived accounting quantity for stored recursive compression within an active sovereign attractor. The durable imprint is the α-trace (ALPHA, § 8.3): the persistent structural record left by repeated recursive selection. § 8.3 gives, as a canonical cross-substrate example, exactly this domain: “trained weights, sparsity patterns, attention biases — what a neural network has learned is the α-trace of its training history.”
Kernel mapping
| Observable in a memory system | Kernel construct | Canon |
|---|
| Path-non-neutrality of recall / replay loops | T0-χ Path-Closure Sensitivity | § 1.3 |
| Episodes → few stored patterns (the compression) | RC recurcline Rc(t) | § 8.2 |
| Durable synaptic / weight imprint | ALPHA α-trace α(X,t) | § 8.3 |
| History needs path-dep + identity + feedback | THT (history ⇐ T0-χ∧T0-B∧T0-M) | § 3.12 |
| Stored pattern only readable on recall | BROP basin-relative observability | § 3.11 |
| “Memory exists only when active” vs “trace at rest” | BAEP regime distinction | § 3.10 |
The two constraints the kernel imposes
BROP — observable only on recall
An attractor's organizing invariants are observable primarily along trajectories that remain within its basin. For a memory attractor, the compressed structure is readable only by evoking it — driving the network back into the basin (recall, cued reactivation, replay). A purely passive external recording that never enters the basin samples only transient state.
The only way to observe the compression is to exercise the very path-closure sensitivity (T0-χ) that made it possible. The closed loop is simultaneously the storage mechanism and the read channel.
BAEP — active attractor vs inert α-trace
§ 8.3 defines two α-trace modes: live α-trace, which “actively filters or biases incoming states” (a memory being recalled), and inert α-trace, the static residue that “does not actively shape future dynamics” (the potentiated synapse at rest). The kernel is explicit: “Sovereignty requires recursion lock; α-trace alone is insufficient.”
A potentiated synapse at rest is inert α-trace, not an active attractor. “LTP is the memory” names the trace, not the attractor — the attractor exists only while the loop runs.
Necessity vs sufficiency
Necessity holds — and is close to definitional. Wherever genuine recursive compression occurs, path-closure non-neutrality (T0-χ) is underneath it: a system whose loops carried no invariant effect would leave nothing for recursion to fold and keep. This is argued from the structure, not established by survey — so the load-bearing content is not necessity itself, but the two claims that can actually fail: sufficiency (below) and observability (above).
Sufficiency fails — by design. Compression-as-recurcline is a co-presence consequence: it needs T0-χ (loop non-neutrality) and T0-M (recursion that revisits the loop) and T0-B (keeping the compressed structure from diffusing), under ignition (IST, § 3.2).
Control case — hysteresis. A magnetic or plastic-deformation hysteresis loop has T0-χ (closed loops carry invariant effect) but lacks T0-M self-reference. Result: loop memory without recurcline. This is why path-closure capacity is a genuine dividing line — mere hysteresis on one side, compressive recursion on the other. (The word “holonomic” is avoided here on purpose: the kernel renamed the v1 primitive “χ-Holonomy” to Path-Closure Sensitivity precisely to stop it smuggling in a manifold the substrate does not have.)
Predictions with testable edges
- No passive readout of the invariant. You cannot recover the stored pattern from a purely passive/external measurement that never drives the system into the basin (
BROP). Readout requires re-traversal.
- Trace-not-attractor at rest. Between recalls the “memory” is inert α-trace: structural residue present, but no active recursion running. The observable that stands in for
Rc(t) > 0 is simply whether the system is being actively driven in its basin or sitting at rest — see the note on measurement below.
- Compression requires self-reference, not just loops. A system with closed-loop hysteresis but no
T0-M will show loop memory but not episode→pattern compression. Hysteretic materials are the control.
- New trace requires a live loop. Fresh α-trace is laid down only while the recursion is running and maintaining itself; once active maintenance has collapsed, a system cannot consolidate new memory even though its accumulated trace persists. Stated in observable terms (not a measured
Rc value): where ongoing active maintenance fails, new-memory consolidation should fail while old memory remains.
On measuring recurcline. One caution the framework insists on: recurcline is estimated, not defined — the construct's own text says “metrics estimate Rc(t); no metric defines it.” So the predictions above ride on an observable proxy — whether the recursive loop is actively running or at rest — not on a measured recurcline value, which does not yet exist. Turning Rc into a real quantity is open, candidate work (V1–V2), and the promising direction keeps two things apart: the cost of holding the compression (an accumulated maintenance-action taken step-by-step along the sequence, which corresponds substrate-by-substrate to energy expenditure, without being it) and the amount compressed (an information-structural measure such as statistical complexity). Until that lands, read every Rc on this page as “active recursion present,” not as a number.
Connection to reserved kernel work
The kernel reserves, as future domain-scoped Tier-1 work (§ 13+), a Hopfield bootstrapping-interval theorem — named alongside the Bio-Kernel derivations. This note is upstream scaffolding for that reserved theorem, not a claim that the theorem exists. The bootstrapping interval I_B (§ 8.2) — over which Rc(t) and α(X,t) are defined, bounded below by ignition and above by β-loss — would map to the interval over which a trained network actively maintains its stored attractors.
Boundary Notice
This page is downstream, candidate structural correspondence (V1–V2), not empirical alignment. It applies canonical constraints; it does not modify, reinterpret, or extend them. Conflicts are resolved in favor of the canonical text and governance pages. To move any claim here toward stronger evidence classes, it must be held against real source material.