The cognitive layer for AI agents
MokenLogic is an AI systems lab working on persistent memory, structured reasoning, and learning from experience. Our systems combine foundation models with durable state and inspectable structure to study how agents remember, reason, and adapt over time.
Working thesis / 001
The next step in capable agents is not only a larger context window. It is a durable cognitive architecture that can preserve experience, recover structure, test conclusions, and improve across tasks.
Research programs
We treat an agent as a changing system—not a stateless call to a model. Our work concentrates on the structures that carry useful experience from one decision to the next.
Experience as a queryable state
We study memory systems that preserve episodes, entities, relations, and provenance across long-running agent work—then recover only what the current decision needs.
Structure before fluent answers
We explore representations that make relations, constraints, and multi-step conclusions explicit enough to inspect, compose, and test.
Open-source repositoryMeasured updates from experience
We investigate how agents can improve from tasks and feedback without turning every interaction into an opaque retraining event.
System architecture / model-agnostic
MKL—SYS—01
Foundation models provide powerful priors. The MokenLogic layer is designed to coordinate durable state, relational structure, feedback, and runtime tools around the model for long-lived work.
Not another prompt chain.
Not retrieval without state.
Not learning without evidence.
System in practice / 001
The MokenLogic system is designed to carry a task from intent to a verified deliverable, while preserving the evidence and state needed to continue later.
Working patterns
03 / patternsCollect sources, preserve the claims they support, assemble a brief, and leave the next session with a traceable starting state.
Source-backed briefRun a repeated monitor or report, compare the new result with prior state, and ask for attention only at an exception or decision point.
Scheduled reportResume a long-lived project with its relevant files, decisions, tools, and open questions—without rebuilding the whole context by hand.
Continuable workspaceOpen model systems / Research direction
We are exploring reference systems that pair open-weight language models with MokenLogic memory, tensor-reasoning, and evaluation harnesses. The aim is a reproducible model system whose behavior can be studied across the whole cognitive stack.
The planned release format includes base-model provenance, cognitive-layer versions, evaluation protocols, and known limits.
Inference / Planned API access
Planned API access will expose versioned inference sessions with MokenLogic memory, structured reasoning, and traceable updates around the selected foundation model.
Memory can persist across authorized calls while retaining its provenance.
Cognitive state remains separate from the underlying model choice.
Requests, retrieved context, and updates are designed to remain inspectable.
MokenLogic Harness / macOS
In developmentA native macOS harness for agents working across apps, files, conversations, and long-lived projects. It brings MokenLogic memory, structured reasoning, and inspectable execution into real workflows—where behavior can be evaluated beyond benchmarks.
Designed for macOS.
Explore the HarnessSelected writing / Wayland Zhang
Three independent essays that provide mathematical and systems context for the questions MokenLogic studies.
Four mathematical lenses on neural networks: dynamical systems, manifold geometry, gauge structure, and measurement.
Read essayHow a question in differential geometry reappears in numerical mechanics and learned neural representations.
Read essayA systems design for preserving durable progress, liveness, and recovery inside long-running agent turns.
Read essayLab notes
We separate working systems, active experiments, and longer-term directions. Research language should never outrun the artifact.
MokenLogic / Research systems