AI预检检查：智能体工作记忆架构

In short : A working memory architecture for AI agents : preflight retrieves the right skill from a long-term library, a local model executes, & a watchdog reads the trail overnight to update the library.

I still remember when my agent would forget what I said mid-sentence.

Context size is not the ceiling. Memory architecture is.

Diagram of the memory system : a user request flows through a preflight check that loads the right skill from a long-term skills library into the context window, executed by the local Ornith 35B model, with a watchdog reading the trail at the end.

I have been experimenting with a memory architecture that runs preflight instructions. A pilot plans the route before takeoff. My agent does the same.

A query lands. “Summarize the Q3 board deck.” 200,000 raw tokens of emails, PDFs, & chats sit behind that sentence.

Preflight is retrieval. The agent inspects its skills library¹, picks the ones relevant to the task, & loads only those into the context window. Skills are consolidated memory ; the preflight step is how the agent picks the right one.

The local Ornith 35B model² then executes on that loaded context. Hard tasks route out to the frontier ; routine tasks remain on the local model, which happens about 80% of the time.

Editorial illustration of an airline pilot in the cockpit running a preflight checklist, with a planned flight route curving through the windshield.

The watchdog monitors which skills are loaded, which decisions are made, & the success rate. Every preflight decision is logged. Every skill invocation is a named, versioned artifact.

Overnight, asynchronous inference³ processes the day’s trail. It decides which new skills should be developed, & which parts of existing skills should become deterministic code. Calendar scheduling is a good example : an LLM should not be comparing free & busy slots ; Rust is much better at that. The system rewrites its skills library & restarts itself in a self-improving loop.

Yesterday was the first day the watchdog did not suggest any improvements. I doubt it will continue. But it hints at something : at some level of improvement, the system reaches a plateau. Only genuinely new exceptions need human help.

The skills library is a set of workflow files (~90 at present) indexed on-disk & retrieved by intent match. Skills are workflows written once, versioned, & handed to the model as tool schemas. See Skill Distillation for how the library was built. ↩︎
Ornith 35B is a locally-hosted open-weight model in the 35-billion-parameter class, run on Apple Silicon via Ollama. It handles routine agent work — classification, drafting, tool selection, structured extraction — & routes the hard remainder to the frontier. ↩︎
See Full Sail on Asynchronous Inference for the queue architecture that makes overnight, hours-long agent runs tractable. ↩︎