Skip to main content

ctx-tree

Claude Code with a photographic memory.

ctx-tree is a persistent SQLite property graph that sits between Claude Code and every tool call. When Claude reads a file, runs a grep, fetches a URL, or calls any MCP tool, a hook intercepts the result, stores it in the graph, and returns a compact nodeId reference (~60 tokens) instead of dumping raw data into the context window.

The same operation next week costs zero tokens. It's already in the graph.

The problem​

LLM context windows are a scarce, write-once resource. Tool calls flood the window with low-signal data:

  • Read on a 500-line file: ~4,200 tokens, used once, never freed
  • Grep across 8 files: ~8,400 tokens, immediately stale
  • WebFetch on API docs: ~11,000 tokens, 90% noise

By the time a complex task is 60% done, the window is full of content that already answered its question. Claude starts dropping earlier context. Sessions die before work is done.

This is context rot.

The solution​

ctx-tree intercepts every tool call via Claude Code hooks:

  1. The native tool is blocked — no raw data enters the context window
  2. The equivalent ctx-tree MCP tool runs instead
  3. The result is stored in the graph — chunked, indexed, linked
  4. Claude gets back a compact nodeId (~60 tokens)
  5. Any future session can recall it for free via ctx_tree_compose

The graph persists across sessions. Knowledge accumulates. Context stays lean.

Next steps​