Two things arrived this week that belong in the same sentence and never will be. Estefania Ajcip's sculptural paintings, built from the fractured memories of families separated across borders, and the PrimeAgentOrchestrator paper from arXiv, which addresses how LLM coding agents start each session with an empty context window, discarding everything they previously knew. Both are studies in what it costs to lose continuity. Neither discipline has fully solved it.
The Empty Context Window and the Separated Family
The PrimeAgentOrchestrator paper's premise is blunt: AI agents begin every session with amnesia. The researchers propose memory-primed spawning as a solution, essentially pre-loading agents with structured prior context so each session does not begin from zero. Ajcip's paintings perform the same operation by different means. The folded, layered surfaces are archives: compressed records of what was shared before separation that the family now uses to reconstruct a coherence that geography has dissolved. The formal parallel is not decorative. Both systems are trying to solve for persistent identity across interrupted sessions. The family does it with objects and images. The AI does it with vector embeddings. The question neither fully answers is what gets lost in the compression.
What Coherence Requires
A 2024 paper in Memory and Cognition by Sekeres et al. found that autobiographical memory reconstruction is fundamentally social: we remember better in the presence of others who share the context. The separated family in Ajcip's work is not just missing people. It is missing the social infrastructure that makes memory accurate. The PrimeAgentOrchestrator's memory-priming faces an analogous problem: context pre-loaded without the interaction that generated it is a summary, not a memory. The SDAD paper on spec-driven agentic development gestures at this too: the problem is not storage, it is the lived specificity that context windows cannot fully capture. Eugenia Kuyda, speaking with Culture Slop about Replika, made the same point from a product angle: software shaped by ongoing interaction is categorically different from software pre-loaded with a static profile of a person. Both Ajcip and the AI researchers are building containers for what continuity feels like. Neither is building continuity itself.