Agent Re-Read Reduction Graph

Cache what your coding agent has already read to cut repeated file reads and tokens

Problem

Coding agents reread the same files over and over on long tasks, burning tokens and slowing responses. A developer already built a knowledge graph to reduce re-reads, showing this is a widely recognized waste in agent workflows.

Solution

Maintain a lightweight local knowledge graph of which files an agent read, which snippets and conclusions it reached, expose it via MCP or hooks so the agent queries summaries instead of rereading source files on the next task, with manual refresh for stale nodes.

Users

Developers heavily using Claude Code, Cursor, Codex and similar coding agents

MVP

Local file index and summary generation; MCP service exposing query APIs; integration with at least one mainstream coding agent; automatic invalidation of stale files; a simple hit-rate and token-savings panel

Revenue

Open-source core plus subscription cloud graph sync; or per-seat team subscription

Why now

Coding agents are moving into long tasks and multi-person workflows where context cost becomes a visible expense. The MCP ecosystem is young, so a vertical cache layer can become a default component early.