5
Engine families
Cognitive · Active research
An architecture,
not an end-to-end model.
A non-transformer cognitive architecture organized into five engine families.
1
Shared memory interface
N
Independent teams
What Penelope is
An architecture, not an end-to-end model
Penelope treats cognition the way an operating system treats compute: as a set of cooperating engines with explicit interfaces. Perception, memory, reasoning, safety, and control are five distinct layers, each developed against its own test bed, and composed through a shared structured-memory interface. The architecture is what we ship; the model is what we compose.
Engine families
Five engine families, one architecture
Penelope is organized into five engine families. Each is developed independently and composed through a shared structured-memory interface.
Control
Orchestration · scheduling · resource-aware execution
Safety
Admission · audit · capability boundaries
Reasoning
Analogy · counterfactual · multi-step inference
Memory
Structured binding · substrate-native recall
Perception
Signal → structured evidence
Engine deep-dive
Five engines, in detail.
PP
Perception
Engines that turn raw signal into structured evidence the rest of the stack can reason over.
PM
Memory
Engines built around structured binding and substrate-native recall, not token windows.
PR
Reasoning
Engines for analogy, counterfactual, and multi-step inference over composed memory.
PS
Safety
Engines that gate, audit, and constrain what the rest of the architecture is allowed to do.
PT
Control
Engines that orchestrate the others — planning, scheduling, and resource-aware execution.
Architecture maturity
How the engine families took shape
Foundations
Five families, named and scoped
The five-engine split is committed to as a research contract. Interfaces between families are versioned so each engine can iterate without breaking the others.
Memory first
Memory family wired to Heddle
The memory engine standardises on structured binding. Every other family writes and reads through that interface.
Safety horizontal
Safety engines gate the others
Safety is not a wrapper around outputs; it is an engine family with admission control, audit, and capability boundaries applied at the architecture layer.
Composition
Cross-engine evaluations running
End-to-end tests measure how well the five families compose. Reasoning quality is graded against the rest of the architecture, not in isolation.
What Penelope enables
A platform, not a product
Composed flagship lines
RL-X1 and RL-C1 are composed within Penelope, with each model line picking a specific configuration of the five families.
Architectural alignment
Capability boundaries, oversight surfaces, and refusal behaviour are properties of the architecture, not of a final-layer prompt.
Independent engine research
Each family can be advanced by a different team against the same interface, so the architecture absorbs progress without rewrites.
Related research
The work behind Penelope.
Available through