Graphicle
Code graph and memory for your coding agents, in one Rust program. It maps your repository, remembers your project's docs, decisions and past sessions, and returns cited excerpts.
Explore GraphicleAI research & development
Language models
that remember
millions of tokens.
CIEL, our research model, was trained only on texts of up to 8,000 tokens, yet recalls facts hidden in texts of millions of tokens. We also build open-source tools: Graphicle for coding agents and JLA for linear attention in JAX.
What we build
Two open-source tools, one research program.
Graphicle gives coding agents memory today: a code graph and your project's docs, decisions and past sessions, on your machine. JLA brings fast linear-attention kernels and a training stack to JAX.
CIEL is our foundational research: long-term memory that lets a small language model recall facts far beyond the length of text it was trained on.
JLA
Fast linear-attention kernels for JAX on NVIDIA Hopper: Mamba-2 and Mamba-3, Gated DeltaNet, KDA, RWKV-7 and GLA, with a full training stack. It trains a 1.3B transformer at 72.3% MFU in FP8 on one H100.
Explore JLACIEL
Long-term memory for language models. In synthetic recall tests, our 1B-parameter model, trained only on texts of up to 8,000 tokens, recalls 93% of facts hidden in a 2-million-token text and 91% at 4 million, on one H100.
Explore CIELGet startedWork with
Work with
us.
Run our model on your own GPU: we're looking for design partners. Install Graphicle and your coding agents get its code graph and memory today.