page:docs-supermemory-research-raw-11-smfs-benchmarks
SMFS: Making Agentic Retrieval 55% Cheaper and More Accurate reference
Source: https://blog.supermemory.ai/smfs-making-agentic-retrieval-55-cheaper-and-more-accurate/
SMFS: Making Agentic Retrieval 55% Cheaper and More Accurate
Source: https://blog.supermemory.ai/smfs-making-agentic-retrieval-55-cheaper-and-more-accurate/
Overview
SMFS.ai (Supermemory Filesystem) is a purpose-built filesystem designed for AI agents. Combines agentic search with semantic retrieval to optimize cost and accuracy.
Core Features
- FUSE-powered filesystem with instant loading
- Auto-generated profiles (
/profile.md) that update dynamically - Multi-modal support via OCR (images to searchable text)
- Enhanced grep: semantic search alongside traditional string matching
The Problem
**Agentic search** provides control and structure but struggles at scale -- agents must manually traverse directories and maintain context across operations.
**Semantic RAG retrieval** efficiently finds content but strips context -- returns isolated chunks without surrounding information or file relationships.
Developers were forced to choose between control (agentic) or reach (semantic).
The Solution: xAFS Benchmark
Created a realistic evaluation framework featuring:
- Mixed conversational and document data
- Scalable file counts up to 10,000
- Multi-hop and temporal reasoning queries
- Files exceeding 10,000 tokens each
Performance Results
- **Accuracy**: At 10,000 files, SMFS maintained 81% accuracy vs 69% for baseline filesystems
- **Cost reduction**: 55% cheaper overall ($946 vs $2,103 across evaluations)
- **Token efficiency**: 53.8% fewer tokens used; 53.1% fewer per correct answer
- **Per-query savings**: One corpus showed $4.71 cost vs $20.95 for baseline
- **Claude specifically**: -66% tokens, -60% tool calls with improved accuracy
Technical Approach
Hybrid methodology: 1. Semantic search lands on specific file paths 2. Agent-controlled navigation through surrounding context 3. Targeted grep operations within identified subtrees
Agents trust their starting points while maintaining control over exploration.