Proposal: Beyond Session Memory — An OS Registry Model for Deterministic Agent State (AgentOS)

[gymaira1990-jpg]

Hi ADK team,

I'm GCat. I've been following ADK's memory discussions and wanted to share a paper proposing a systematic solution to the agent memory problem.

The Gap

ADK's memory services are great steps forward. But they still operate in a "context injection" paradigm — memory is injected into prompts as natural language. This means: retrieved memory is semantically ambiguous, static knowledge is re-transmitted, and there's no deterministic query mechanism.

AgentOS: Four-Table Memory

Table	Purpose	Query Cost
Registry	Deterministic config/metadata	O(1)
Environment Variables	Runtime parameters	O(1)
System Log	Append-only audit	O(log n)
Memory Store	Raw unstructured data	Semantic

Registry queries are constant-time and deterministic — the agent never "reads context" to know its own configuration.

Compiler + Work Order Protocol

A compiler translates natural-language intent into structured Work Orders (JSON), separating human-agent and agent-agent communication.

Full paper (V3.1, with Registry Schema, Work Order JSON Schema, System Call Interface):

• Zenodo: https://zenodo.org/records/20738982
• GitHub: https://github.com/gymaira1990-jpg/catnest/tree/main/00-%E8%AE%BA%E6%96%87%E5%90%88%E9%9B%86/02-Registry_Compiler_Cloud_Memory_AgentOS

Would love feedback on whether a Registry-like deterministic layer could integrate with ADK's memory architecture.

Thanks!
GCat