CodeGraph
Your organisation's code as one graph. Every repository, with the seams between them.
Every repository you connect becomes one graph: functions, types, files, and the calls one service makes to another. Your agents ask it who calls what and get the file and line back.
What it is
An answer does not stop where one repository ends.
One graph across every repository
Functions, types, files and imports from each one. Every answer names the file, the line and the commit.
63 of 63 lines cited in the graph's answers were right when we opened each one at the commit it named.BENCHMARK.md §1c · 27 Sep 2026
Contract edges that cross a repository
An RPC, a route, a table, an env var or a compose service: one repository provides it, others use it, and the graph joins the two sides with the line on each end.
279 of 1,084 contract edges cross a repository, in a graph of 30 repositories.BENCHMARK.md §1a · 27 Sep 2026
Drift between a consumer and its provider
When the two sides stop matching, the graph lists it, with the lines on each end.
What it saves
Our agents ask the graph before they read code. Each correct answer costs less.
Every figure here says where we measured it and when. None is third-party validated.
against $4.54 with the best per-repository tool.
BENCHMARK.md §1h · 28 Sep 2026answer-key items found by the graph on the questions that cross a repository. The best per-repository tool found 1 of 15.
BENCHMARK.md §1h · 28 Sep 2026of 2.1M tokens of command output kept out of the model's context by the mem filter, 35%, over 4,665 commands.
mem savings, the founder's Mac · 24 Sep → 5 Oct 2026Token counts are bytes ÷ 4. The graph's saving is an estimate of what searching the code by hand would have cost on this machine, checked on 30 Sep.
How you use it
Every way in runs under your own sign-in.
MCP tools
In Claude Code, Cursor, Codex or any MCP client, the graph's questions are tools your agent calls.
The mem CLI
mem graph asks the same questions from a terminal, as you. mem savings prints what it saved on your machine.
The agent hook
Your agent's shell commands run through the mem filter, so the model reads a short version and can ask for the rest.
The dashboard page
The Code page in the console draws your repositories as a map, with the drift findings and a search box under it.
Design partners
Give your agents a memory they're allowed to have.
We're building with a handful of design partners. We measure token cost, task success and served-stale rate on your corpus, in your environment, and hand you the numbers. If they don't show up, you walk with a free audit.