Skip to main content
Core is REI’s continuing neuro-symbolic learning and reasoning substrate, where declared public task semantics, live evidence, and revisable state meet. A Core instance begins without accumulated local history and changes through use: observations can form task-relevant structure, repeated support can stabilize it, corrections can revise it, weak support can fade, and later work begins from what the instance has already formed. A response is a snapshot of Core’s activity, not its definition. Core uses what it has formed to interpret new work, compare possibilities, expose support and conflict, and revise what later work begins from. Language can sit at a boundary to encode an interaction or articulate a result, but it does not do the continuing reasoning or own the state.

​
The external contract

Public task boundary
Context entering the instance
Declared semantics, interactions, events, constraints, relations, or outcome signals
Reasoning substrate
Core
A continuing substrate where declared meaning and live evidence form and revise task structure
Application boundary
Current task view
A decision or interpretation, its support and uncertainty, and any relevant state change

This is the public contract around Core, not a sequence of internal modules. The current task view reflects both the declared task frame and the evidence-shaped state of the instance.

​
What changes through use

  • A fresh instance begins without accumulated local history.
  • Observations, outcomes, and corrections can form or revise task-relevant structure during operation.
  • Repeated support can stabilize structure while weak support loses influence and multiple supported associations can remain active together.
  • Available product surfaces can expose a current result, state, and supporting evidence.
This describes observable behavior rather than internal implementation details.

​
Configured explicitly, formed through use

Core can begin operating without a task-specific dataset, labeled demonstrations, solved trajectories, or an offline task-training phase. This does not mean zero configuration: the public task interface may still be declared explicitly, while task-local structure forms from live evidence. A fresh Core instance begins without accumulated local task history, not without an interface. The surrounding application or domain should declare the meanings that are public parts of that interface: observation fields, available operations, action semantics, native outcome types, and the expected result shape. Core should still have to learn the episode-specific structure: which visible object has which role, which action works in the current context, which sequence reaches the goal, and which reusable mechanism the evidence supports. Withholding public action or outcome semantics from Core while supplying them to another agent changes the information condition. It does not isolate reasoning quality.

​
Two public surfaces

Both are interfaces to Core. They differ in how an application or person meets the substrate, not in who owns the continuing reasoning or state. Persistent state belongs to the selected Core instance, not to the surface. Direct use also avoids making language quality, token use, or articulation latency part of the reasoning path.

​
Persistent, associative, and revisable

Persistence includes more than retained content. Several supported associations can remain active together, repeated support can make useful structure more influential, and later evidence can change what carries support. This is useful only while state remains correctable. Evidence inspection, explicit correction, isolation, reset controls, and failure-aware operation and validation are therefore part of the product contract.

​
Choose a surface

Build with Adapt-1

Bring Core’s continuing reasoning into an application without a conversational layer.

Explore Rei Chat

Use the conversational demo to experience continuity and revision through language.