Erevu
AI that grows, remembers and learns.
Erevu is a brain-inspired, post-transformer AI architecture designed to grow its capacity, retain a living memory and learn continuously during use. It separates internal reasoning from output generation, with governance built into how the system evaluates meaning and decides when to act.
Continuity, consistency and control.
Inconsistent decisions, lost institutional knowledge, fragmented governance and high operational overhead can make AI difficult to use in critical business processes. Erevu brings memory, learning and governance into the architecture, with controls designed to check for contradictions, identify uncertainty and govern when output is committed.
The Inner Voice
The Inner Voice is Erevu’s internal governance layer. It is designed to evaluate what the system is doing and why, identify when more information is needed, detect instability, constrain risky behavior and recognize when a task is complete.
Its role is to govern the system’s behavior and output commitment, separate from generating conversational text.
Learning that accumulates
Erevu is designed to learn incrementally through plasticity and consolidation, with explicit controls to preserve stability.
- Strengthen: reinforce connections that are used frequently.
- Weaken: let idle connections decay and prune those that no longer contribute.
- Develop: add structure in areas with sustained activity.
- Simplify: reduce low-activity structure to limit noise and drift.
Knowledge that stays available
A persistent memory system is designed to store and retrieve knowledge while protecting it as the model learns and grows. Working memory is designed to help preserve context, while long-term memory supports recall, audits and controlled updates.
The aim is to keep useful knowledge available across interactions and reduce context-loss errors.
Progress beyond the prompt
When there is no external task, Erevu’s architecture is designed to consolidate memories, resolve unfinished threads and identify knowledge gaps through a rest-state cognition loop.
Persistent internal state supports continuity across interactions, with learning and governance operating over time.
Research milestones
Our lab work has demonstrated organic growth across four consecutive phases and reached near-production-quality output at one billion parameters.
- Lab-validated scale
- 1 billionparameters in the research model
- Organic growth
- 50M to 1Bacross four consecutive growth phases
- Training data
- ~1/145of industry-standard data in the reported research
The growth phases were completed without catastrophic forgetting. A separate resilience comparison, validated at 50 million parameters, reported 400× greater resilience than standard AI of comparable size. These figures describe the stated lab models and research stages.
Preparing for first production deployment.
Erevu remains an active research and development architecture and is not yet a generally available commercial product. Full-budget training and alignment are underway to prepare the architecture for its first production deployment.
We welcome conversations with organizations interested in early collaboration.
Discuss early collaborationThe development journey
Explore the timeline: 2023 to 2026
- August 2023
Research began into governed, continuous intelligence.
- Late 2023
Early prototypes separated cognition, meaning and output.
- Early 2024
Continuous governance was introduced through the Inner Voice.
- Mid 2024
Plasticity-style updates and separate memory subsystems were prototyped.
- Late 2024
Internal builds added idle-time cognition loops and bounded curiosity.
- Early 2025
Activity-driven growth and pruning mechanisms were enabled.
- Mid 2025
System-wide governance, diagnostics and safety controls were strengthened.
- Late 2025
A single model grew from 50 million to 1 billion parameters across four consecutive phases.
- Early 2026
The one-billion-parameter model reached near-production-quality output using approximately 1/145 of industry-standard training data.
- Next
Complete full-budget training and alignment for the first production deployment.
Frequently asked questions
Is Erevu available commercially today?
Erevu is in active research and development and is not yet generally available. Organizations interested in early collaboration can contact Infogility to discuss opportunities.
How does Erevu differ from a Transformer-based LLM?
Erevu’s design separates internal reasoning from output generation and combines persistent memory with continuous governance. The distinction is architectural: it is designed to preserve state, learn incrementally and control when output is committed over extended periods of operation.
What does continuous operation mean?
The architecture is designed to maintain internal state across interactions without session resets, retaining context and learning incrementally from ongoing experience. Between tasks, it is designed to consolidate memories and identify gaps through its rest-state cognition loop.
How do memory and governance work together?
Persistent memory provides the knowledge and context the system can retrieve. The Inner Voice monitors and constrains its behavior, checking uncertainty and stability and governing output commitment. Together, these mechanisms are designed to support auditable reasoning and controlled operation as the architecture grows.