{
  "id": "page:docs-supermemory-research-raw-11-smfs-benchmarks",
  "_kind": "Page",
  "_file": "wiki/docs/supermemory-research/raw/11-smfs-benchmarks.md",
  "_cluster": "wiki",
  "attributes": {
    "nodeKind": "Page",
    "sourcePath": "docs/supermemory-research/raw/11-smfs-benchmarks.md",
    "sourceKind": "repo-docs",
    "title": "SMFS: Making Agentic Retrieval 55% Cheaper and More Accurate",
    "displayName": "SMFS: Making Agentic Retrieval 55% Cheaper and More Accurate",
    "slug": "docs/supermemory-research/raw/11-smfs-benchmarks",
    "articlePath": "wiki/docs/supermemory-research/raw/11-smfs-benchmarks.md",
    "article": "\n# SMFS: Making Agentic Retrieval 55% Cheaper and More Accurate\n\nSource: https://blog.supermemory.ai/smfs-making-agentic-retrieval-55-cheaper-and-more-accurate/\n\n## Overview\n\nSMFS.ai (Supermemory Filesystem) is a purpose-built filesystem designed for AI agents. Combines agentic search with semantic retrieval to optimize cost and accuracy.\n\n## Core Features\n\n- FUSE-powered filesystem with instant loading\n- Auto-generated profiles (`/profile.md`) that update dynamically\n- Multi-modal support via OCR (images to searchable text)\n- Enhanced grep: semantic search alongside traditional string matching\n\n## The Problem\n\n**Agentic search** provides control and structure but struggles at scale -- agents must manually traverse directories and maintain context across operations.\n\n**Semantic RAG retrieval** efficiently finds content but strips context -- returns isolated chunks without surrounding information or file relationships.\n\nDevelopers were forced to choose between control (agentic) or reach (semantic).\n\n## The Solution: xAFS Benchmark\n\nCreated a realistic evaluation framework featuring:\n- Mixed conversational and document data\n- Scalable file counts up to 10,000\n- Multi-hop and temporal reasoning queries\n- Files exceeding 10,000 tokens each\n\n## Performance Results\n\n- **Accuracy**: At 10,000 files, SMFS maintained 81% accuracy vs 69% for baseline filesystems\n- **Cost reduction**: 55% cheaper overall ($946 vs $2,103 across evaluations)\n- **Token efficiency**: 53.8% fewer tokens used; 53.1% fewer per correct answer\n- **Per-query savings**: One corpus showed $4.71 cost vs $20.95 for baseline\n- **Claude specifically**: -66% tokens, -60% tool calls with improved accuracy\n\n## Technical Approach\n\nHybrid methodology:\n1. Semantic search lands on specific file paths\n2. Agent-controlled navigation through surrounding context\n3. Targeted grep operations within identified subtrees\n\nAgents trust their starting points while maintaining control over exploration.\n",
    "documents": []
  },
  "outgoingEdges": [],
  "incomingEdges": [
    {
      "from": "page:docs-supermemory-research",
      "to": "page:docs-supermemory-research-raw-11-smfs-benchmarks",
      "kind": "contains_page"
    }
  ]
}