Cognitive runtime
Cognitive runtime
Section titled “Cognitive runtime”Chat works well for short interactions. Longer jobs are harder: they may need to survive a restart, remember what happened earlier, recover from failure, wait for a person, or switch models without losing the task itself.
The Cognitive Runtime is the layer LM Nexus is building for that kind of work.
Main parts
Section titled “Main parts”The responsibilities are deliberately split up:
- Durable Work Runtime — keeps goals, plans, tasks, execution attempts, events, outputs, checkpoints, and task state.
- Memory — stores useful remembered information together with where it came from and what scope it belongs to.
- Context Builder — decides which relevant pieces are actually sent to the model for a given request.
- Model Router — chooses a compatible provider/model for the work instead of assuming one fixed default.
- Agent Runtime bridge — connects durable task state to Agent Runtime execution and resume.
- Runtime Supervisor — decides what to retry, recover, or escalate when work stalls or fails.
- Human Tasks — represents an explicit pause when the system needs a person to do or decide something.
They cooperate, but they do not share one giant “AI state” database.
Preview maturity
Section titled “Preview maturity”| Area | What works in the current integration | What still needs to land |
|---|---|---|
| Durable Work Runtime | focused integration tests are passing | final integration plus broader local/final checks |
| Memory + Context Builder | backend, UI, and context-focused tests are passing | final integration; Dreaming/Reconciliation remains later work |
| Model Router core | core routing stages are passing focused tests | Resource Coordinator still needs a public Model Manager interface for preparing/checking runtimes |
| Agent Runtime bridge | restart-safe bridge is passing focused tests | still depends on final integration of the full stack |
| Runtime Supervisor | recovery through human escalation is passing focused tests | Router fallback still needs a durable record of why a route was chosen |
| Human Tasks / delegation foundation | child success/failure paths are passing focused tests | tracked delegation still waits on reliable verification and parked-parent coordination |
What Preview means here
Section titled “What Preview means here”Preview means the implementation exists and important focused tests are passing on the current integration work.
It does not mean:
- every component is already merged to the stable default branch;
- every UI/runtime consumer is integrated;
- every native, browser, or hardware gate is closed;
- the final public API or schema is frozen;
- the first public release guarantees these contracts unchanged.