Resonant Attractor Memory as a Portable, Compressed Context Layer for Frozen Language Models

Transformer language models store their working context in a key–value (KV) cache that grows linearly with sequence length and can exceed the size of the model weights. Current remedies — eviction, quantisation, and low-rank compression — treat context as a sequence to be pruned. This position paper argues for a different primitive: storing accumulated context as a fixed-point attractor in a bolt-on associative-memory module attached to a frozen LLM, and recalling it by relaxation rather than sequential reconstruction. The proposal rests on mature attractor theory and on recent (2025–2026) work on compressed continuous-time Hopfield memories, yet occupies an unfilled gap: no existing system places an attractor memory as a compression layer between a running transformer and its context, and none exploits the resulting representation as a portable, model-independent memory artefact that can migrate across devices and model generations. The paper motivates the idea, positions it against the active KV-compression literature, sketches a falsifiable single-GPU pilot, and treats the portability–privacy tension as a first-class design concern. This is an idea/position draft, not a peer-reviewed result.