The mental model shift from agents that “forget everything when the conversation ends” to “agent as operating system” reframes every architectural decision. (Note: this OS metaphor is an editorial synthesis of Rohit’s memory architecture work — his original content focuses on practical memory patterns rather than using the OS analogy explicitly.) Conversation context is RAM — fast and volatile, lost on reset. Persistent memory is the hard drive — indexed, structured, survives sessions. Memory decay is garbage collection — without it, the system drowns in stale data. Tool integrations are I/O management — how the agent reads from and writes to the external world.
This framing explains why The context window is the fundamental constraint — everything else follows feels so fundamental: it’s a RAM limitation, and every workaround (subagents, specs files, /clear between tasks) is a memory management strategy. The OS model also clarifies why Compound engineering makes each unit of work improve all future work works: CLAUDE.md files, skill docs, and pattern libraries are the “installed software” that makes the OS more capable over time, persisting across reboots. And just as operating systems maintain crash dumps and error logs, Agents that store error patterns learn continuously without fine-tuning or retraining by storing failure patterns as a distinct memory tier — the agent equivalent of /var/log. It’s the architectural basis for The three-layer AI stack: Memory, Search, Reasoning — Memory, Search, and Reasoning are the OS kernel services that applications build on.