BENCHMARK · MEASURED, NOT CLAIMED

Does it actually help?

On questions about a team's own codebase, Claude Code was measured with and without TeamGraph. The memory made the agent faster, cheaper, and markedly more likely to be right — and it stopped inventing answers.

Reproducible On a team's own code Run it yourself
teamgraph · benchmark results CLAUDE CODE
Benchmark: with TeamGraph, Claude Code answers 40% faster with 49% fewer tokens, right answers go from 27% to 73%, and made-up answers from 27% to 0%

Questions about a team's own codebase, answered by Claude Code with and without TeamGraph.

27→73%Right answers about the team's own code
27→0%Answers that were made up
49%Fewer tokens on solved problems
40%Faster, end to end
WHAT THE NUMBERS MEAN

Right more often.
And honest when it isn't.

It stops making things up

Without TeamGraph, a quarter of the answers about the team's own code were invented. With TeamGraph, none — the verify gate returns a MISS instead of a confident guess.

A wrong note doesn't derail it

On 30 questions no note answers, TeamGraph sometimes offered the wrong note — and the agent set it aside every time. Bad recall costs nothing; it just isn't used.

Where it still slips

TeamGraph finds the right note for 88% of questions. The misses are usually a question about the same technology as another note — the next thing being worked on.

BROKEN DOWN

Fewer tokens, better recall.

tokens & speed
Chart: TeamGraph reduces tokens and time on solved problems
recall quality
Chart: TeamGraph finds the right note for 88% of questions
The reproducible benchmark — how it was measured, and how to run it yourself — ships in the repository's bench/ directory.
MEASURE IT ON YOUR OWN CODE

Trust the number you ran yourself.

Install TeamGraph, point it at your project, and see the difference on the questions your team actually asks.