Why One Search Algorithm Is Never Enough: A Taxonomy of AI Memory Retrieval
BM25, HNSW, GraphRAG, temporal decay - a practical breakdown of the search paradigms powering AI memory systems, organized by the failure mode each one solves.
-
-
Prompt Caching: The 30% You're Leaving on the Table
Most developers optimize model choice, context size, and output quality, but never notice the 20-35% of their bill that disappears with one config flag. Here's what prompt caching actually is, how five major providers implement it differently, and what happens when your tooling silently breaks it.
-
I Benchmarked Every Embedding Model Worth Running. Here's What I'd Actually Deploy.
We went from nomic-embed-text to OpenAI's text-embedding-3-small and thought we'd upgraded. Turns out we'd moved from bad to mediocre. Here's the full landscape of self-hosted embedding models in 2026, organized by what you can actually run on your hardware.
-
Bigger Context Windows Don't Make Better AI
The arms race to 1M context windows assumes more context is better. Research from Chroma proved it makes things worse - and labs have been expanding context windows ever since.