FS-Researcher: почему вашему ИИ пора завести блокнот и перестать тупить

FS-Researcher: Why Your AI Should Get a Notebook and Stop Being Dumb

Imagine you're trying to write a doctoral dissertation, but you have the memory of a goldfish. You read the fiftieth source, and at that moment, the details of the first simply evaporate. This is exactly how even the most advanced language models like GPT-4o or Claude 3.

5 Sonnet feel right now. They have a context window that is growing, but the "Lost in the Middle" problem hasn't gone away. When there's too much data, the model starts to get confused, ignoring important details from the middle of the text and producing superficial mush instead of deep analysis.

This is a classic case where quantity doesn't translate into quality, and engineers have spent years trying to solve this by simply expanding working memory.

A group of researchers decided to approach the question differently and presented FS-Researcher. Instead of jamming hundreds of pages of search results into an unfortunate model, they gave it a cognitive prosthesis in the form of a structured notebook. The idea is simple to the point of genius: the agent doesn't just search and read, it actively takes notes. In the process of working, the system highlights key entities, facts, and connections, writing them to an external storage that is constantly updated. This allows the model to maintain focus on the task without overwhelming the main context with garbage that inevitably appears during deep web searches.

Previously, we relied on RAG—a system that pulls chunks of text from a database. But RAG often works like a bad librarian: it brings the right page but doesn't understand the big picture. FS-Researcher works like a thoughtful analyst. It hierarchically organizes information, filters out duplicates, and most importantly, knows how to connect facts found in different sources. If one document talks about the cause of an event and another about its consequences, the system doesn't just copy both paragraphs, but synthesizes them into a single logical chain in its notebook. This is critically important for writing long reports, where not only factuality matters, but also the structure of the narrative.

Why is this important right now? We've hit the ceiling of "brute force" context windows. Companies like Google brag about millions of tokens in Gemini, but in practice, using such a huge window costs an insane amount of money and slows down generation to snail's pace. FS-Researcher shows that architectural decisions and proper agent workflow are much more efficient than endless parameter expansion. For business, this means that quality automated analysis of markets or technology trends becomes cheaper and more accurate. You no longer need to check every fact the neural network produces, because it itself maintains a transparent log of its reasoning and findings.

Ultimately, we are observing a transition from AI-chats to AI-employees. A chat just answers a question, but an employee is someone who can work with information for a long time, systematically, and not forget why they opened the browser ten minutes ago. FS-Researcher is one of the first confident steps toward creating autonomous researchers who actually save human time instead of adding work by correcting errors. This changes the game in academia and the corporate sector, where data accuracy always comes before speed of delivery. Now the question is just how quickly such overlays become the standard for all popular LLM services.

The main point: the era of "bloated" context windows may end before it even begins, giving way to smart filtering systems and external notebooks. Will OpenAI and Anthropic be able to implement such mechanisms natively in their models, or are we in for a boom of third-party agent platforms?