Fraction — AI skill for Claude Code

Fraction

Persistent memory layer for LLM agents and AI applications. **3x faster** than LLM-based extraction — zero API costs, sub-100ms ingestion, fully offline.

Outperforms **mem0** on BLEU-1 (+17%), F1 (+10%), and LLM Judge (+14%). Outperforms **supermemory** on BLEU-1 (+8%) and LLM Judge (+5%). All with **zero LLM calls** and **1.6x faster ingestion** than mem0.

Fraction supports two extraction modes:

• LLMLingua-2 (default) — learned token compression, zero API cost, fully offline
• LLM extraction — any LLM provider via litellm (OpenAI, Anthropic, Ollama, etc.)

Both modes use the same hybrid retrieval layer: vector similarity + BM25 + entity graph + temporal boost, merged via Reciprocal Rank Fusion.

Installation

pip install fractionally

# Download the spaCy model for entity extraction
python -m spacy download en_core_web_sm

# Optional: install litellm for LLM-based extraction (supports any provider)
pip install fractionally[llm]

Quick Start

from fraction import Memory

m = Memory()

# Add memories
m.add("I love hiking in the Rocky Mountains.", user_id="alice")
m.add("My favorite book is Dune by Frank Herbert.", user_id="alice")
m.add("I'm allergic to peanuts.", user_id="alice")

# Search memories
results = m.search("outdoor activities", user_id="alice")
for r in results["results"]:
    print(f"{r['memory']} (score: {r['score']:.3f})")

# Memories auto-persist to ~/.fraction/

Features

• Two extraction modes — LLMLingua-2 (free, offline) or LLM-based (any provider via litellm)
• Zero API cost (default mode) — compression + embedding + retrieval run locally
• Sub-100ms ingestion — LLMLingua-2 compression + USearch indexing
• Deterministic — same input always produces same memory (LLMLingua mode)
• Hybrid retrieval — vector similarity + BM25 keywords + entity graph, merged via Reciprocal Rank Fusion
• Auto-persistence — memories survive process res