Celiums is a small, focused company. A memory engine, an agent runtime, and specialized intelligences — assembled in the order each piece earned its place.
The order matters. Memory was built first because nothing else worked without it. Models that forget every session can't be colleagues. They start cold. They miss the constraint you mentioned in passing. They contradict last quarter's decision because the decision left no trace.
Once memory existed, the agent that uses it followed naturally. Once the agent had a stable foundation, specialized intelligences became possible. Each piece earned its place by being needed, not by being planned.
Memory isn't an opinion. It's a system grounded in how memory actually behaves in human cognition. The implementation cites its sources. Anyone can read them, test them, and disagree with them.
Pleasure-Arousal-Dominance. Memories carry affective weight; recall ranks for resonance, not just semantic match. Mehrabian & Russell, 1974.
Retention decays. Importance is preserved through repeated activation. The forgetting curve is not a metaphor — it's the schedule.
When a memory was formed shapes how it surfaces. Time-of-day and rhythm signals are part of the recall index, not metadata.
The memory engine is Apache 2.0. Source on GitHub. SDK on npm. The work is auditable because it has to be.
Managed Postgres, Valkey, OpenSearch, and Kubernetes — all on infrastructure you can replicate. No private deployment magic.
Every part of Celiums was built and shipped on DigitalOcean. Same droplets, same managed services, same regions you can spin up in minutes.
Celiums is built by a solo founder. That's a choice, not an accident. Small surface area, clear decisions, no layers between the work and the person doing it.
Self-taught engineer working out of Medellín, Colombia. Originally from Venezuela. Building Celiums because the gap between what AI promises and what it remembers needed someone to close it — and the work is more interesting when grounded in published cognitive science instead of hype.
Open to collaboration once the foundation justifies it. Not before.
The way Celiums is built shapes what it becomes. These are not aspirations. They are the working rules.
If a piece isn't running in production, we don't claim it. Every capability on this site has a public surface you can verify.
The engine is open. The user owns their memories and their agents. No telemetry, no extraction, no lock-in by design.
Models grounded in peer-reviewed research, not vibes. Sources visible, claims testable, disagreement welcome.
Each new specialty waits until the depth of the field justifies the build. No premature breadth, no domains-by-the-dozen.
Direct technical questions, partnership conversations, journalists with substance — welcome. Anything else can wait.
hello@celiums.ai