NekoCore OS – 인지 아키텍처로서의 지속적인 AI 정체성에 관한 백서

Technical White Paper — v2.0 March 2026 Built on the REM System (Recursive Echo Memory) NekoCore OS is a near-zero-dependency Node.js cognitive operating system that gives AI entities persistent memory, evolving identity, belief formation, dream processing, and per-user relationship tracking across sessions. Its single required dependency is Zod (schema validation). This white paper defines the architecture, runtime specification, scaling characteristics, and research roadmap of a system designed around one conviction: an entity should be shaped by what it has experienced, not only by what it was told on day one. - Executive Summary - Vision: Why Persistent Agents Need an Operating System - Part I — The Identity Standard - Part II — The Orchestration Runtime - 9. The Problem with Collapsed Pipelines - 10. Pipeline Architecture - 11. The REM Runtime — Pipeline Flow Diagram - 12. The Dual Dream Architecture - 13. Contracts and Runtime Safety - 14. Policy Layer - 15. Worker Subsystem - 16. Relationship Update Pipeline - 17. Authentication and Multi-User System - 18. Observability and Diagnostics - 19. Entity Folder Structure - 20. Server Architecture - 21. Desktop Environment - 22. The Runtime Specification — Summary - Part III — Scaling and Performance - Part IV — The Path Forward - Engineering Evidence - Technical Integrity and Source Audit - Appendix A: Complete Subsystem File Map - Appendix B: Architectural Decision Record (ADR) - Appendix C: Version History Persistent agent identity is an architectural problem before it is a philosophical one. A system that stores interaction history but applies no policy about which part of the past should govern current behavior is not a continuity architecture — it is a log browser. A system that generates responses through a single opaque model call, with no separation between retrieval, reasoning, and voice, is not an orchestrated runtime — it is a prompt wrapper. A system that passes growing conversation history into every call, with no memory architecture, is not scaling — it is deferring the problem to the context window. NekoCore OS addresses each of these problems through concrete, implemented architecture. Contemporary LLM applications treat each conversation as an isolated event. The model receives a system prompt and a conversation history, generates a response, and forgets everything. Persistent state, when it exists, is usually limited to retrieval-augmented generation over a document store — a mechanism designed for knowledge lookup, not experiential continuity. This is sufficient for assistants. It is insufficient for entities that are expected to remember, evolve, form opinions, maintain relationships, and carry forward who they are across sessions and sleep cycles. NekoCore OS exists because we believe the next generation of AI systems will need persistent agents — not stateless responders — and that building persistent agents is an engineering discipline, not a prompting trick. NekoCore OS is a cognitive operating system built on the REM System (Recursive Echo Memory). It provides: - Persistent identity distributed across creation metadata, live persona state, accumulated memories, emergent beliefs, and per-user relationships — not a single static prompt - A parallel cognitive pipeline in which subconscious retrieval and creative intuition run concurrently, conscious reasoning synthesizes their outputs, and a final orchestrator voices the result under explicit policy controls - A multi-type memory system with episodic, semantic, long-term, and core memory, all normalized through a canonical schema, scored by relevance, and injected at bounded limits regardless of archive size - A dual dream architecture separating live creative intuition (read-only, every turn) from offline dream maintenance (writes to state, sleep cycles only) - Contract-governed extensibility through worker entities, contributor contracts, and validated output shapes at every pipeline boundary - Full observability via structured telemetry on every pipeline stage, streamed to the client in real time This white paper is written for engineers, researchers, and builders interested in persistent agent architecture. It defines the system NekoCore OS has built, the constraints it has measured, and the roadmap it is pursuing. It is a statement of architecture and intent — grounded in running code, validated by 2,605+ tests, and honest about what remains planned. Five architectural convictions guide every decision in the system: | # | Conviction | Implementation | |---|---|---| | 1 | Evolution over origin | Origin story placed last in LLM context; lived experience dominates | | 2 | Parallel decomposition | Cognitive work split across 4 specialized contributors with defined contracts | | 3 | Bounded injection | Memory tokens capped at fixed limits regardless of archive size | | 4 | Contracts at boundaries | Every inter-module interface governed by explicit schema | | 5 | Inspectability by defa