Opportunity: Codebase Knowledge Layer for Localz

The match

Projects/Localz.md is explicitly a synthesis of 39 raw doc files, compiled specifically to give coding agents (and Mihir) "a single reference to drive concrete implementation decisions" — i.e. Localz already hit the exact problem this window's memory/knowledge-layer coverage addresses: raw docs are too scattered for an agent to re-derive architecture from cold, every session.

Two tools this window describe the productized version of what the Localz synthesis report does by hand:

• Concho AI — pre-analyzes a codebase into a queryable knowledge graph once ("cognitive precompiler"), exposed to existing coding assistants via MCP instead of re-deriving architecture from raw files every session.
• Tools/Graphify — the open-source, self-hosted, no-account version of the same idea: tree-sitter builds a deterministic AST-based knowledge graph locally, exposed as an MCP server with 10 query tools, integrates directly with Claude Code.

Relevance

Right now the Localz synthesis report is a manually-maintained markdown file playing this role. That's fine pre-implementation, but once actual FastAPI/Next.js code exists, the same "re-derive everything from raw files every session" cost this window's coverage describes will start applying to the real codebase, not just the docs. Graphify in particular is a low-cost trial candidate — self-hosted, MIT, works with Claude Code directly, no account required — worth a quick spin (graphify install + /graphify .) once there's enough real code in the repo to be worth mapping, rather than defaulting straight to grep-based exploration as the codebase grows.

Not yet actionable

This is a "worth trying once the codebase exists" note, not an immediate task — Localz is still pre-implementation per its own synthesis report status.

Related

• Projects/Localz
• Synthesis/agent-memory-latency-and-generalization — the broader synthesis this specific match is drawn from