# KarLex AI > KarLex AI builds Graphicle, open-source code graph and memory for coding agents; JLA, open-source linear-attention kernels for JAX; and CIEL, research on long-term memory for language models. KarLex AI, Inc. (also written KarLexAI) was founded in 2024 by Vikram Karlex R. Its work spans agent memory, machine learning, quantum computing and verifiable computation. ## Projects - [Graphicle](https://www.karlex.ai/graphicle): Graphicle maps your repository into a code graph, remembers your project's docs, decisions and past agent sessions, and gives coding agents cited excerpts. One Rust program, open source under Apache-2.0. Version 0.1.10, released 2026-10-30; source code at [graphicle-ai/graphicle](https://github.com/graphicle-ai/graphicle). - [JLA](https://www.karlex.ai/jla): JLA (jax-linear-attention) brings FLA-class linear-attention kernels to JAX: Pallas Mosaic GPU kernels for NVIDIA Hopper, forward and backward, for Mamba-2, Mamba-3, Gated DeltaNet, KDA, RWKV-7 and GLA, with a full LLM training stack. Open source under Apache-2.0. Version 0.1.0, released 2026-10-21; source code at [jla-org/jax-linear-attention](https://github.com/jla-org/jax-linear-attention). - [CIEL](https://www.karlex.ai/ciel): CIEL is KarLex AI's research on long-term memory for language models. In synthetic recall tests, our 350M-parameter model recalls facts from documents up to 1 million tokens long, after training only on documents of up to 8,000 tokens. A research preview. ## Research - [All papers](https://www.karlex.ai/research): KarLex AI's papers on Transformer learning, quantum computing, and verifiable search, with visual summaries and journal status. - [Local Credit Assignment for CPU Transformers: Readout Consensus and the Cost of Predictive Coding](https://www.karlex.ai/research/bounded-horizon-local-transformer-training-cpus): What happens when a Transformer learns in separate stages? Two CPU studies examine the speed, memory, and learning quality of local training and predictive coding. Under review · Machine Learning. - [Exact softmax sampling from residual quantum overlaps](https://www.karlex.ai/research/exact-softmax-residual-overlaps): Compute part of an attention score classically, then sample the remainder. This study shows how the split changes measurement cost while preserving the ideal softmax distribution. Under review · Quantum Information Processing. - [Shot saturation and evidence limits in quantum-AI hardware benchmarks](https://www.karlex.ai/research/shot-saturation-quantum-ai): Do more quantum measurements produce a better answer? A reanalysis of hardware records tracks the error that remains as measurement budgets grow. Under review · Discover Quantum Science. - [Proof-Carrying Optimality for Finite Identification under Bounded Adversarial Answer Errors](https://www.karlex.ai/research/proof-carrying-finite-identification): Find the fewest questions needed to identify an answer, even when some replies are wrong. Then give an independent checker a proof that the question count is optimal. Under review · Theory of Computing Systems. - [Quantum Hybrid Modules for AI: Attention, Optimization, and Verification on Near-Term Quantum Hardware](https://www.karlex.ai/research/qhm-quantum-hybrid-modules): The original QHM project explores quantum subroutines for AI. Its hardware records and attention questions lead into two focused follow-up studies. Preprint · Research Square. ## Graphicle comparisons - [All comparisons](https://www.karlex.ai/graphicle/vs): How Graphicle, open-source code graph and memory for coding agents, compares with Graphify, codebase-memory-mcp, CodeGraph, Mem0, Hindsight, claude-mem, Graphiti and Cognee: where each one fits, what Graphicle adapted from them, and what we measured. - [Graphicle vs Graphify](https://www.karlex.ai/graphicle/vs/graphify): Graphicle's code-graph engine is a Rust port of Graphify, credited in its NOTICE, with Graphify's MCP tool names. It adds project memory, session capture and an app, in one local program. - [Graphicle vs codebase-memory-mcp](https://www.karlex.ai/graphicle/vs/codebase-memory-mcp): codebase-memory-mcp and Graphicle both give coding agents a code graph over MCP from a native program, with docs linked to the code and no API key. Graphicle also captures past agent sessions and keeps memories you approve. - [Graphicle vs CodeGraph](https://www.karlex.ai/graphicle/vs/codegraph): CodeGraph and Graphicle both build a local code graph that coding agents query over MCP, with no API key. Graphicle also keeps your project's docs, reviewed decisions and past sessions, and collects no telemetry. - [Graphicle vs Mem0](https://www.karlex.ai/graphicle/vs/mem0): Mem0 is a memory layer for AI apps and agents. Graphicle remembers a repository's docs, decisions and past agent sessions for coding agents, beside a code graph, in one local program with no account. - [Graphicle vs Hindsight](https://www.karlex.ai/graphicle/vs/hindsight): Hindsight is an open-source memory system for AI agents. Graphicle adapts parts of its design and keeps a repository's docs, decisions and past agent sessions beside a code graph, in one local program. - [Graphicle vs claude-mem](https://www.karlex.ai/graphicle/vs/claude-mem): claude-mem gives Claude Code a memory of past sessions. Graphicle's session capture follows its design, and adds a code graph, your docs and reviewed decisions. - [Graphicle vs Graphiti](https://www.karlex.ai/graphicle/vs/graphiti): Graphiti, from Zep, builds temporal knowledge graphs for agent memory in a graph database. Graphicle adapts parts of it and keeps a repository's memory and code graph in one SQLite file. - [Graphicle vs Cognee](https://www.karlex.ai/graphicle/vs/cognee): Cognee turns documents, data and code into a knowledge graph for AI agents. Graphicle remembers a repository's docs, decisions and past agent sessions, beside a code graph, for coding agents. ## Company - [About](https://www.karlex.ai/about): Learn about KarLex AI, the company behind Graphicle, JLA and the CIEL memory research, with work in machine learning, quantum computing, and verifiable computation. - [Contact](https://www.karlex.ai/contact): info@karlexai.com - Elsewhere: [GitHub](https://github.com/graphicle-ai), [Hugging Face](https://huggingface.co/KarLexAI), [LinkedIn](https://linkedin.com/company/karlexai), [X](https://twitter.com/karlexai), [Discord](https://discord.com/invite/rWuU7cWU4E), [Instagram](https://www.instagram.com/karlexai) ## Discontinued - ASIOS, an AI-native operating system built on Ubuntu, was an exploratory project KarLex AI announced on Linux kernel mailing lists in May 2025. It has since been discontinued: ASIOS is not a KarLex AI product and is not in development. KarLex AI's current work is Graphicle, JLA, CIEL and its research papers. ## Optional - [Terms of service](https://www.karlex.ai/legal/terms) - [Privacy policy](https://www.karlex.ai/legal/privacy)