Answers are grounded in this paper and cite their sources.
A transformer keeps its working memory of a conversation in its key–value (KV) cache, and throws it away when the session ends. Leviathan, a local inference engine for consumer NVIDIA GPUs, keeps it. The KV cache is organised as a multi-resolution *fractal cascade*: the most recent tokens are held exactly, and older ones are folded into progressively coarser summaries (a centroid plus its principal directions per group). The whole cascade is written to disk as a `.fqm` file and restored when the model next loads. This revision describes the system as it now stands: a CUDA implementation; atomic, verified writes; a selectable storage stage (half-precision by default, about 2× smaller at 0.02 % error); turn-boundary compaction so files plateau instead of growing; a structural health check; and a way to read a memory back, which turns out to matter for privacy. It also reports what the memory does and does not keep. Restored memory carries the gist, tone and continuity of past sessions. Specific facts that have been folded into the compressed levels do not survive, even with lossless storage. Exact recall belongs in text, and Leviathan pairs the `.fqm` with a text layer for it.
No takes yet. Share an insight, caveat, or question.
Michael Ricky Neal (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: