Eternis develops self-evolving AI systems focused on forecasting and simulating uncertain futures using calibrated predictions. The company offers products like Freysa for agent simulation and Silo for private, locally-hosted model access. Their work centers on maintaining continuously updated world models and enabling epistemic search capabilities for AI agents.
Funding
Funding not disclosed
Founders
Product
Problem
Current digital twin solutions often lack continuous learning capabilities, robust memory architectures, and verifiable security, limiting their ability to accurately represent individuals and act on their behalf with trustworthiness. Existing systems struggle to adapt to new information, maintain consistent personal histories, and ensure secure interactions with external tools.
Solution
Eternis is developing infrastructure for building capable AI digital twins that faithfully mirror an individual's personality and can represent them accurately at scale. Their research focuses on continuous learning, enabling digital twins to evolve with every interaction through various reward regimes. They are also creating a multi-layered memory architecture that allows twins to ingest, store, and reason over a lifetime of memories, ensuring consistency and a rich personal history. To address security concerns, Eternis is building verifiable Model Context Protocol (MCP) servers within Trusted Execution Environments (TEEs), providing mathematical verification of proper behavior and preventing credential theft, tampering, and tool poisoning attacks.
Target Audience
Eternis targets developers and organizations building AI-powered digital twins and related applications, particularly those requiring continuous learning, robust memory, and verifiable security.
Features
- Continuous learning pipeline with fast in-session adaptation using Adapter/LoRA heads for real-time refinement.
- LLM-judge preference loops for batch fuzzy-reward data and offline RLHF with calibrated LLM evaluators.
- Hierarchical, multi-timescale RL combining daily proxies with monthly targets under a multiscale controller.
- Multi-layered memory framework with multiple query types, including local, global, and episodic query support.
- Iterative data ingestion mechanism for rapidly ingesting new data without restructuring existing memory.
- Proactive memory application that volunteers relevant past information at contextually appropriate moments.
- MCP servers deployed within Trusted Execution Environments (TEEs) for verifiable execution guarantees.
- Cryptographic attestations proving that only approved, unmodified code is running.
- Protection against tool poisoning attacks by ensuring tool descriptions match exactly what was approved and published.