fs-researcher-scaling

name	fs-researcher-scaling
title	FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.01566
keywords	["Test-Time Scaling","External Memory","Multi-Agent","Long-Horizon Reasoning","Knowledge Accumulation"]
description	Scale research agent capability using persistent filesystem as external memory. Dual-agent architecture with context builder accumulating knowledge and report writer composing outputs enables computation scaling beyond context windows.

FS-Researcher: Test-Time Scaling via External Memory

Problem

Research tasks require accumulating knowledge far exceeding model context windows. Deep research needs iterative refinement, synthesis across sources, and integration with external findings.

Single-pass generation cannot match quality of human iterative research. Standard scaling approaches hit context limits.

Core Concept

FS-Researcher uses the filesystem as durable external memory and coordination medium. A Context Builder agent accumulates knowledge by web browsing and extracting information; a Report Writer composes final output using the persistent knowledge base.

Allocating more computation rounds to the Context Builder directly improves downstream report quality, demonstrating effective test-time scaling.

Architecture Overview

• Context Builder Agent: Web browsing, information extraction, hierarchical knowledge base construction
• Report Writer Agent: Section-by-section composition using persistent knowledge base
• Filesystem Coordination: Shared external memory enabling cross-session persistence
• Hierarchical Storage: Organize findings into nested structures
• Iterative Refinement: Multiple rounds of context building improve knowledge base quality
• Section-by-Section Writing: Incremental report composition with knowledge base lookups

Implementation

Step 1: Build Knowledge Base Storage

Create hierarchical filesystem structure for knowledge accumulation.

import os
import json
from pathlib import Path

class KnowledgeBaseStorage:
    def __init__(self, base_path):
        self.base_path = Path(base_path)
        self.base_path.mkdir(parents=True, exist_ok=True)

    def add_finding(self, topic, finding_text, source_url, metadata=None):
        """Store a research finding."""
        topic_path = .base_path / topic.replace(, )
        topic_path.mkdir(parents=, exist_ok=)

        
        finding = {
            : finding_text,
            : source_url,
            : datetime.now().isoformat(),
            : metadata  {}
        }

        
        findings_file = topic_path / 
         (findings_file, )  f:
            f.write(json.dumps(finding) + )

     ():
        
        topic_path = .base_path / topic.replace(, )
        findings_file = topic_path / 

          findings_file.exists():
             []

        results = []
         (findings_file, )  f:
             line  f:
                finding = json.loads(line)
                 query.lower()  finding[].lower():
                    results.append(finding)

         results

     ():
        
        topic_path = .base_path / topic.replace(, )
        findings_file = topic_path / 

          findings_file.exists():
             []

        findings = []
         (findings_file, )  f:
             line  f:
                findings.append(json.loads(line))

         findings