Claim
I don't trust working memory for anything important. Every customer interview, every competitive datapoint, every claim I might one day put in a deck — it goes into a file with a category, a date, and a source. Then synthesis is a query, not a recollection. The wiki, the codex you're reading, every research repo I run — same pattern, different scopes.
Mechanism
Memory decays. Files don't. When the substrate is queryable, the synthesis layer becomes mechanical, which means an LLM can do most of it and you spend your judgment on the edges. Without substrate, every synthesis is a from-scratch recall, which is expensive and wrong.