时间观察者 · timeobserver137

I build multi-agent systems that remember.

An eight-agent fleet with a memory architecture, retrieval pipelines, and this hand-built site — plus a shelf of long-form essays on symmetry, group theory, and Euler.

View projects Read writing GitHub
L1 L2 L3 L4 L5
Effective-potential contours of the restricted three-body problem, L1–L5 — Euler found the collinear three first. From my essay on Euler's blind years.

The language of symmetry

The universe allows exactly five perfect solids — their symmetries are page one of the Symmetry & Monster series, a road that ends at a beast with a 54-digit number of them. Start from episode one →

Featured work

All 7 →

OpenClaw — an eight-agent fleet that remembers

featured

Eight persona agents with a shared memory architecture, running my research, engineering, and media production for months on end.

LLM agentsPythonRAG / embeddingsDiscord buscron / systemdgit-as-memory

Meinrag — a multimodal RAG platform

featured

Document intelligence for PDFs that fight back: layout-aware parsing, hybrid retrieval with reranking, and an MCP layer so agents can use it as a tool.

PythonFastAPIPostgreSQLFAISSDoclingMCP

Meinopoly — turn any book into a playable world

featured

A moddable Monopoly-style strategy engine: one command and an AI pipeline reads a novel into characters, boards, portraits and balance-tuned rules; LLM agents even take real seats at the table over MCP. 1,489 unit tests + 45 E2E.

JavaScriptboardgame.ioLLM pipeline (extract/build/art/balance)MCP serverJest (1,489 tests)Playwright E2E

Latest writing

All essays →

对称与怪兽（六）·潘多拉的魔盒

中文

对称与怪兽（五）：三十年，一万页，与一个没人读得完的证明

中文

给 AI 装一套会遗忘的记忆:一个八智能体系统的记忆架构全史

中文

八个 AI 智能体,一套版本化的记忆系统,和它从 2026 年 2 月至今每一次失败的完整记录。记忆为什么会膨胀、自相矛盾、把有用的东西藏起来——以及我们如何用真实的 git 历史推翻自己的增长预测,给记忆装上一道会遗忘的门。

#AI 合作 #记忆系统 #多智能体