Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Fat-Cat: A document-centric context management Agent. Making context as simple as reading chat history.
| Date | Stars |
|---|---|
| 2026-08-03 | 724 |
| 2026-08-06 | 724 |
| 2026-09-10 | 723 |
| 2026-09-16 | 724 |
| 2026-09-22 | 725 |
| 2026-10-14 | 724 |
| 2026-10-30 | 724 |
| 2026-11-04 | 724 |
Today
— stars today
This week
— stars this week
This month
-1 stars this month
Momentum
0.0
growth rate 0.00%/day
# Fat-Cat: The LLM-Native Operating System A next-generation Agent framework based on global document context and multi-stage reasoning <p align="center"> <img src="image.png" alt="Fat-Cat Framework" width="800"> </p> ## 1. Background & Pain Points: Why Do We Need Fat-Cat? In the current LLM Agent development paradigm, engineers are facing two core challenges: "the quagmire of context management" and "fragile control flow". I summarize these challenges as the three original sins of Agent design: ### Pain Point One: The JSON Trap Traditional Agent frameworks (such as early LangChain patterns or Assistant API) tend to pass state through complex JSON objects or list dictionaries. **Problem:** LLMs are essentially trained on text. Forcing models to parse deeply nested JSON states leads to attention dilution—models often see the trees but miss the forest, easily overlooking critical constraints. **Engineer's Nightmare:** When debugging, facing thousands of lines of JSON dumps makes it difficult to intuitively understand what the Agent is actually "thinking". ### Pain Point Two: The Static Toolset Most Agents have "hard-coded" capabilities. When facing unknown problems, Agents can only operate within preset if-else statements or fixed DAG graphs. They lack the ability for **Runtime Learning** and cannot acquire new skills by consulting resources like humans do. ### Pain Point Three: The Absence of Metacognition This is the most fatal weakness of current Agents—"only execution, no reflection". **Phenomenon:** Traditional Agents receiving tasks act like reckless interns, directly starting to call tools. Once they hit a dead end (such as code errors or failed searches), they often fall into infinite retry loops or generate hallucinations, forcibly providing wrong answers. **Missing Link:** Lack of a high-level "monitor" process to evaluate: "Am I doing this correctly?", "Can my current strategy solve this problem?", "Do I need to stop and replan?". Fat-Cat aims to solve the above problems. It is not just a Bot that executes tasks, but an operating system prototype with "self-awareness" and "evolutionary capabilities". ## 2. Core Design Philosophy ### 2.1 LLM as Operating System In Fat-Cat, we treat LLM as CPU, Context (document context) as memory (RAM), and external tools as peripherals (I/O). The Fat-Cat framework itself acts as the Kernel, responsible for process scheduling (Stage switching), memory management (Memory Bridge), and exception handling (Watcher Agent). ### 2.2 Document as Global Context We abandon fragmented JSON and adopt Markdown documents as carriers of global state. Each Stage's output is a "revision" or "supplement" to this global document. - Stage 1 generates reasoner.md (problem analysis document) - Stage 2 generates strategy.md (tactical manual) - Stage 3 generates step.md (SOP execution table) - Stage 4 executes and backfills results. This design makes the Agent's "thinking process" completely visible and debuggable to humans. ## 3. Core Features Deep Dive: Fat-Cat's Metacognitive System Fat-Cat's core breakthrough lies in constructing a hierarchical metacognitive closed loop. This is not simple Prompt Engineering, but rather forcing Agents to "think twice before acting" through architecture. 🧠 **Stage 1: Metacognitive Analysis (Deep Intent Perception)** "Think about how to do it before starting" Traditional Agents receiving "help me write a crawler" might directly start writing code. But in Fat-Cat, Stage 1 Agent (Metacognitive_Analysis_agnet.py) will force metacognitive analysis through reasoner.md: - **Intent Decomposition:** Does the user really just want code, or do they need deployment? - **Constraint Extraction:** What are the implicit language, performance, and dependency library requirements? - **Information Completeness Check:** If information is insufficient, it will refuse to execute and request suppleme
Excerpt of 12,751 characters
Read on GitHub
61
2
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:d1828685503952b7, llm:Repository description: 'Fat-Cat: A document-centric context management Agent. Making context as simple as reading chat history.' Language: Python. No topics provided.
matched fp:d1828685503952b7, llm:Repository description: 'Fat-Cat: A document-centric context management Agent. Making context as simple as reading chat history.' Language: Python. No topics provided.