How to cut AI coding agent token costs by indexing code outside the repo

A developer writing for DEV Community describes how loading entire monorepos into an AI coding agent's context window was inflating token usage and session costs. The core insight is separating 'domain memory' — notes, contracts, and canonical decisions — from cheap, session-level code lookups such as finding a symbol or its callers. The proposed solution stores a symbol index in a local cache outside the repository and git history, so the agent can query it without re-reading the full working tree each session. The author recommends using only one code-read tool per session, compressing noisy CLI output, and reindexing whenever the agent confidently references symbols that no longer exist. Four checklist questions are suggested before indexing any repo: where the index lives, whether it enters a write-capable agent context, how it can be deleted or revoked, and who else has access to it.