- Author
- KarLex AI, Inc.
- Published
- Undated; benchmarks dated October 2026
- Kind
- GitHub repository README for a developer tool, with extensive self-run benchmarks
- Original
- github.com/graphicle-ai/graphicle
- Subject
- Agent tools and setups
What it is
KarLex AI, Inc. publishes the README. Graphicle builds a code graph and stores docs and past sessions. It returns cited excerpts to agents.
What it says
- The README says Graphicle took 3.4 to 23.9 times less time than Graphify. This held in 34 of 39 measurements.
- In 5 updates with nothing to change, Graphify was faster. The README says it stops early.
- It reports LoCoMo-10 accuracy of 87.7, level with Hindsight and above Mem0 OSS at 81.9.
- It says Graphicle made about 12 model calls per conversation, against about 160 and 332.
- On finding code, it says Graphicle tied grep and does not claim to save tokens.
- It says there is no telemetry. The one automatic request is a daily release check.
- It shows a bug flow: a model reproduces it, finds the code and patches it.
What evidence it gives
The README reports speed and memory on 13 codebases. It adds agent tests, memory comparisons and a retrieval test. In one agent test the agent never called Graphicle's tools. Methods sit in linked pages, which were not read. Chart data was read from image descriptions.
Interests and limits
The company, which has a commercial edition in private preview, wrote and ran every test. The README says the memory results use a pre-release build and development sets. It says long sessions are not yet tested. The linked benchmark pages were not opened. The chart images were not viewed.