Eidos
A user, not a feature.
In DevelopmentAn AI identity standard that lets agents accumulate context, preferences, and trust across sessions instead of starting from zero every time.
What it is
Eidos is a persistent autonomous agent. It has its own identity — real email, real OS user account, real audit trail. It has persistent memory — it learns from every task and becomes better over time. It makes deliberate decisions using structured multi-model reasoning before committing to action. It runs continuously, not as ephemeral sessions.
Most AI agents are stateless tools that forget everything, operate under borrowed credentials, and make decisions in isolation. Eidos is the opposite.
Three pillars
The Discipline
The framework is simple — one core loop, ~500 lines. Simplicity is a structural constraint. The system stays general by refusing to bake in special cases.
The Separation
Purpose flows inward — the human sets the mission. Capability flows outward — the agent executes. The agent doesn't invent its own goals. It executes yours with judgment, and surfaces when it sees a better path.
The Identity
Eidos is a user, not a feature. Like onboarding a real team member: its own email, credentials, filesystem, and killswitch. This eliminates privilege escalation, attribution, recovery, and sandboxing problems.
Pods
Before acting, Eidos convenes three different models:
Memory
PEFM — Poisson Emotional Forgetting Memory. Every memory carries emotional weight from real-world outcomes. Successes reinforce. Failures warn. Unused memories decay on a Poisson curve. The agent doesn't just remember — it remembers what mattered, weighted by how it turned out.
Retrieval is hybrid — BM25 + vector + keyword with RRF fusion — so it finds what matters regardless of how you ask.
Status
Eidos v5 is the current engine. The core loop, Pod deliberation, and PEFM memory are operational. The governance trilogy, agent identity (AID), and EidosOmni memory system are the infrastructure layers being built around it.