Karpathy recently pitched an autonomous LLM-driven personal knowledge base (LLM-Wiki). The internet blindly praised it.

The architecture is naive: dump raw notes into a raw/ folder and let cron-triggered LLMs scan, extract, and recursively link everything into a wiki/ directory. While this plays well in toy environments, executing it in production is an engineering disaster.

The O(N²) Compute Black Hole

The design violates basic computational complexity limits. It works for 10 articles. When the database scales to 10,000 files, adding a single new note triggers a global recalculation. The LLM must align entities and rebuild the network graph across the entire dataset. This is an O(N²) operation. It is not an architecture; it is a mechanism to burn tokens with zero ROI. Production engineering demands strictly bounded resource consumption.

State Pollution and Cascading Failures

The LLM-Wiki assumes models are flawless text processors. They are not. If an LLM hallucinates during an unsupervised cron merge—confusing two variables or merging distinct concepts—the error is hardcoded into the file system.

Subsequent automated runs read this contaminated state as truth, compounding the error. This creates an unrecoverable cascading failure. Without strong state isolation and human gating, the credibility of the database drops to zero.

The Limina Architecture: JIT Injection and Async Dual-Core

Complex systems require state control. At Limina Engineering, we solved this with physical isolation:

1. Cold/Hot Separation via JIT Stop running global updates. Core system context remains static in a hot MEMORY.md. Cold data is never processed proactively; it is fetched Just-In-Time via deterministic, lightning-fast retrieval engines (qmd / mem-search).

2. Asynchronous Dual-Core Pipeline Data ingestion and knowledge synthesis are decoupled.

Ingest (Fetch & Archive): Dumb, local models (oMLX Qwen) run scraper scripts to dump web data into raw/. Zero premium tokens used.
Synthesize (Wiki-Builder): Cron triggers are disabled. When a human architect explicitly commands a build, a strict LangGraph workflow activates. Premium models (Claude 3.5 Sonnet) resolve conflicts and write to wiki/.

Autonomous N-squared LLM loops are architectural laziness masquerading as innovation. Build strict boundaries, isolate state mutations, and keep the execution trigger in human hands.

🇨🇳 中文翻译 (Chinese Translation)

Karpathy 最近提出了一个全自动的 LLM 个人知识库架构（LLM-Wiki），引来全网盲目追捧。

其底层逻辑极其天真：将零碎笔记扔进 raw/ 目录，靠定时任务触发大模型去扫描、提取并重构网状链接到 wiki/ 目录。这套东西在玩具环境里跑得通，但在真实的生产环境中是彻头彻尾的灾难。

O(N²) 算力黑洞

该设计完全无视了系统复杂度常识。10 篇文章跑全局重构可行；当库里有 10,000 篇文章时，新增一条哪怕极微小的记录，大模型为了对齐实体和重建图谱，理论上都需要执行 O(N²) 级别的全量重算。这不是架构，这是单纯为了烧 API 额度而设计的无底洞。工程底线要求资源消耗必须具备可控边界，毫无 ROI 的算力挥霍不可接受。

状态污染与错误复利

LLM-Wiki 架构建立在一个极其脆弱的假设上：模型绝不犯错。一旦大模型在无人值守的合并中产生幻觉（例如强行缝合两个独立概念），脏数据将被直接固化入文件系统。

随后的自动更新会把这批被污染的数据当作 Ground Truth 继续衍生，形成致命的错误复利（Cascading Failures）。缺乏强状态隔离和人类干预节点，整个知识库的信任度会迅速崩溃，且根本无法回滚。

Limina 架构：JIT 注入与异步双核

复杂系统的核心是状态控制。Limina Engineering 采用物理隔离彻底重构了该链路：

1. 冷热分离与 JIT 按需检索 停止无意义的全局刷新。高频系统上下文死锁在极简的热缓存（MEMORY.md）中。冷数据彻底放弃主动处理，仅在查询触发时，依靠确定性的检索引擎（qmd / mem-search）执行毫秒级的 JIT（Just-In-Time）精准拉取。

2. 异步双核流（Async Dual-Core） 强制将“数据摄入”与“知识整合”物理断开。

摄入层（Fetch & Archive）： 极度廉价的本地模型（oMLX Qwen）配合脚本无脑抓取数据并落盘至 raw/。严禁在此环节浪费高级 Token。
整合层（Wiki-Builder）： 彻底废除定时触发器。仅在人类架构师明确下达指令时，启动由 LangGraph 强编排的 Workflow，调用高级模型（Claude 3.5 Sonnet）执行冲突解决并写入 wiki/。

迷信“全自动 O(N²) 更新”只是架构设计上的懒惰。建立绝对的边界，隔离状态变更，并把执行的扳机牢牢握在人类手里，才是真正的工程学。