Problem
Agents need persistent identity, long-term memory, and capability discovery to operate across sessions and contexts. Without an OS-level abstraction, each agent reimplements these primitives poorly.
Approach
aiOS provides kernel-level services for agents:
- persistent agent identity with cryptographic attestation
- structured memory with configurable retention and retrieval
- capability registry for tool discovery and permission management
- inter-agent communication via typed message channels
An agent registers with the kernel, receives an attested identity, declares its capabilities (e.g. "can query SQL", "can write files"), and requests access to shared resources through typed syscall-like interfaces.
Architecture overview
The kernel exposes a minimal API modeled after OS syscalls. Memory is layered into three tiers: working memory (ephemeral, cleared per task), episodic memory (session-scoped, persists across tool calls within a conversation), and semantic memory (persistent, survives across sessions and feeds long-term learning).
The capability registry ensures agents only access tools they've been granted. When Symphony orchestrates a multi-agent workflow, aiOS handles identity resolution and memory scoping per agent. When the Control Metalayer enforces a policy, aiOS provides the identity context for per-agent budget tracking.
Current status
Active development with identity, memory subsystem, and capability registry in early iterations. Integration with Symphony (via symphony-aios crate) and Control Metalayer is underway.
Why it matters
aiOS is the foundation layer that turns isolated agents into a composable ecosystem with shared infrastructure.