K-Lines: A Cognitively-Grounded Four-Memory Architecture for Persistent Conversational AI

We present K-Lines, a cognitively-grounded architecture for persistent conversational AI that organizes long-term memory into four complementary types-episodic, semantic, procedural, and reflective-drawing on Minsky's K-line theory, Tulving's memory taxonomy, Anderson's declarative-procedural distinction, and research on sleep-dependent memory consolidation. The central theoretical claim is that the persistent memory problem in conversational AI is not merely an engineering challenge but a cognitive architecture problem-one that Minsky anticipated decades ago, yet that the current wave of agentic AI systems largely reinvents without acknowledging. Effective long-term memory requires the same functional differentiation that cognitive science has identified in biological memory systems. Episodic memory preserves the temporal fabric of conversational interactions. Semantic memory abstracts stable knowledge as structured entities and facts. Procedural memory, operationalized as K-line templates, captures reusable response strategies through the mechanism Minsky described: partial activation records of successful cognitive episodes. Reflective memory synthesizes higher-order patterns from accumulated knowledge. A consolidation pipeline implements computational analogs of sleep-dependent memory consolidation, transforming labile episodic traces into stable semantic representations and synthesizing reflective insights-mirroring the hippocampal-to-neocortical transfer observed in biological systems. All memory artifacts are stored as human-readable files, preserving the interpretability that opaque embedding-based approaches sacrifice. In an empirical evaluation across eight experiments with synthetic domains, the system achieves 65% average cross-session fact recall-a +53 percentage-point improvement over the 12% context-window-only baseline-with per-fixture recall ranging from 29% to 91% depending on domain difficulty. Critically, the system maintains 100% distractor rejection (versus 34% baseline), never fabricating facts that were not taught, and achieves response quality that outperforms both the context-window baseline and no-retrieval condition. Remaining recall gaps are attributable to extraction model capacity rather than architectural limitations, suggesting a clear scaling path.