Introduction to Kyros
Kyros is an open-source, biologically-inspired persistent memory operating system designed specifically for AI agents. Traditional large language models are stateless; they forget conversation context, factual attributes, and procedural steps between interactions. Kyros bridges this gap by offering secure, self-correcting, and audit-ready memory management.
Memory Typologies
Episodic Memory
Tracks chronological conversation streams, tool logs, observations, and raw interactions. Uses vector indices to recall relevant histories.
Semantic Memory
Stores consolidated facts as entities and predicates. Dynamically propagates belief adjustments to resolve contradictory details.
Procedural Memory
Captures step-by-step agent execution workflows. Matches context descriptors against saved procedure templates to guide next steps.
Key Engineering Principles
- Ebbinghaus Temporal Decay: Importance weights degrade over time according to category rates, preventing index bloat.
- Merkle Tree Cryptographic Auditing: Every change creates a hash signature added to a validation ledger to detect poisoning.
- Adaptive Belief Propagation: BFS node traversals recalculate confidences in facts when contradictory statements are ingested.