Faster classification
The Loop’s first stage — Classify — got materially faster this week. The model begins work earlier on simple turns and never blocks on context-fetch for complex ones. The improvement compounds across every interaction with every supported client.
A second client surface
Until this week, gramatr ran in one developer environment. This week it began running in a second — a code-editor client with parity to the original surface. Platform shims for the major LLM client surfaces went in alongside it. The platform stopped being tied to any one harness.
Typed entity taxonomy v1
The first version of the typed entity taxonomy went in mid-week. Decisions, tasks, principles, patterns, runbooks, examples — each with typed enforcement at the write boundary. Every memory entity is structured, scoped, and queryable from this point forward. This is the substrate the audit-trail story rests on.
Better search
Search relevance improved platform-wide in the same window — across every supported surface, with no change visible to the user other than answers getting noticeably better.