Kaipsul is an independent AI research lab. Our current project focuses on context engineering.
Important Notice: MIRC (Research Preview)
Memory-Isolated Recursive Compression ("MIRC") utilizes algorithmic segmentation and probabilistic compression. Output constitutes a lossy representation, not a verbatim reproduction. Users must independently verify all output against original source text. MIRC is provided strictly for research and evaluation purposes. Do not rely on compressed output for legal filings, evidentiary support, or factual verification.
Memory-Isolated Recursive Compression for document pre-processing
MIRC is a document pre-processing tool that segments large texts into memory-isolated chunks, compresses them in parallel using on-device AI, and then reconstructs the compressed output.
It is designed for downstream AI processing. It aims to preserve semantic structure while reducing token count.
Chunking: Segment document into predefined memory chunks
Compression: Process each chunk independently using on-device AI
Reconstruction: Concatenate processed segments into a unified file
Downstream Integration: Output is formatted for LLM inference
Why compression matters for downstream processing
As documents get longer, language models distribute attention across exponentially more tokens. This dilutes attention mechanisms, degrading retrieval accuracy and instruction following.
Many documents often exceed practical limits. Even when technically processable, performance degrades with length.
MIRC increases signal density. By mathematically reducing token count while retaining semantic pointers, it allows downstream AI models to allocate attention resources more effectively.
Designed for AI Systems: Compressed output serves as an intermediary format. Large documents are compressed into a dense format, enabling inference by systems that would otherwise be constrained by context window limits.
Empirical results from MIRC implementation on benchmark documents
Supreme Court Opinion
Federal Legislation
Federal Legislation
Special Counsel Report
Supreme Court Opinion
Reference implementation for macOS
MIT License - Open Source
This is a Swift implementation using Apple's Foundation Models framework. It utilizes on-device processing with parallel compression via actor-based concurrency.
Requirements: Apple Silicon (M1+) - macOS 26.0+ - Apple Intelligence