v3 of the JLC paper series. Theoretical reframing (capacity-not-compression) of the v1 and v2 empirical antecedents archived at Zenodo. This v3 manuscript is a single-file PDF preprint mirror of the arxiv submission (cs.CL). Abstract. Transformers are stateless. That was the power against RNNs — and the curse beyond 30 turns. This paper does not fight the curse — it moves memory to a place the curse cannot reach. Starting from the daily friction a user without ML academic background experienced with an AI companion, a single 1000-turn / 32.62-hour run was completed on commodity hardware: zero infrastructure errors across all turns, 99.25% reduction in cumulative prefill tokens compared to an equivalent dense-context baseline, and mono-lingual verbatim recall at Δ=969 within a 418-token working buffer capacity. After the 1000-turn run, it was confirmed that this architecture maps with 1:1 fidelity to the Complementary Learning Systems framework of McClelland, McNaughton, and O'Reilly (1995); the author was unaware of CLS until after the design was fixed. This correspondence is reported as a suggestion, not claimed as proof. JARVIS — Justified Autonomous Recall via Indexed Silhouettes — is the stateless-native LLM codec proposed in this paper. In long-conversation workloads (turn 30+ / cumulative 50K+ tokens), the inter-turn prefix-carrier role of KV cache accounts for approximately 99% of the compute that caching saves, while the intra-pass autoregressive generation role accounts for approximately 1%. JLC separates the prefix-carrier role at the operational layer and operates through a split of a small sparse buffer (JHB) plus a full-fidelity offline transcript surfaced when needed (JRE): silhouettes are carried quickly, and context fills in the content. Preceding empirical artifacts. v1 — 7 runs × 100 turns, retriever-offline 99.3% semantic recall — archived at 10.5281/zenodo.19681402. v2 — 1000-turn run — archived at 10.5281/zenodo.19776258. Paper thesis: We do not break the curse — we learn to live within it. 24 pages, 0 figures. Source code release (Apache-2.0) pending.
Jun Hyung Kim (Mon,) studied this question.