AI Engineering / Context Engineering / billion-context
LAWSON’S LEARNING NOTEBOOK / SOURCE STUDY 01
A long session.
A small window.
Explore how billion-context folds consumed history into recoverable digests. Read one chapter, change one thing, see what happens.
Begin with active context →Choose a chapter.
Textbook / Reference
Start with the source
Basic idea of LLM context
What the model can actually see
The long-context problem
A growing history, a finite workspace
Core idea
Cumulative work is not resident context
Basic approaches
The common context-management toolkit
The billion-context approach
What distinguishes this particular combination?
Fold
Compress in chunks. Keep the source references.
Hierarchical Compression
Across time. Up through tiers.
Kernel
From raw message identities to compression blocks and the next working view
Compression Doctrine
A prompt-based policy for compression judgment
Growth Gate
When to ask is not what to compress
Recovery
Follow the references back to the source.
ABOUT THIS NOTEBOOK
Based on Lawson Dong’s authored notes and ranxianglei/billion-context ↗, reviewed at 03d27f9 (v0.1.180). Studies use transparent local models. Range selection, source lineage and recovery are grounded in the pinned proxy and kernel; concise notes explain the boundaries.