Home / The second brain

Querying a graph rather than rereading everything

A knowledge graph built once answers a question by walking explicit links, which costs noticeably fewer tokens than a full reread of the source files, a gap illustrated by a public example that carries its own caveat, and the same principle applies to an old session transcript, queried by levels before being reopened in full.

A knowledge graph links the elements of a project, files, functions, notes, through explicit links rather than through their mere position in a folder, and answering a question by walking these links costs noticeably less than a full reread of the source files. Graphify, a third-party tool built for assistants such as Claude, Cursor or Codex, applies this principle to a project's code: it is not an Anthropic feature, but a separate piece of software that builds its own graph once, then answers questions by querying it rather than rereading every file.

A figure, with its own caveat

The most concrete public example comes from Graphify's own blog: a community user reports there having consumed roughly seventy-one point five times fewer tokens by querying their project's graph rather than directly rereading the whole set of their source files for the same question.

The gap is stark, but the article that publishes it describes it itself as a figure relayed by the community rather than a benchmark result, with no published method that would allow it to be reproduced identically on another project, and with no absolute token value attached. Retaining the principle, a queried graph costs less than a full reread, matters more than retaining a precise factor calculated from a single testimonial.

The same principle for an old session transcript

The logic extends to another type of heavy file, the transcript of a long past work session. Rather than reopening a whole transcript every time a question concerns an old decision, a three-level method limits the reading to what is strictly necessary:

level 1: read the summary reminder, a few lines
if the answer is enough, stop here
level 2: grep the relevant passage in the transcript, read only the targeted portion
if the answer is enough, stop here
level 3: read the whole transcript, only if the previous two levels still leave doubt

This sequence limits the reading to what is strictly necessary. A summary reminder first gives a short answer. If that reminder is not enough, a targeted portion of the transcript, around the passage concerned, takes over. Reading the whole transcript stays reserved for the last level, when the note seems out of date or the answer looks doubtful. This method is not a function that Claude offers on its own, it is something you build, like nightly memory consolidation, which can rely on an already refreshed graph rather than rereading everything each night.