ai-persona-os

ai-persona-os is a full operating system for OpenClaw agents rather than a single feature. It ships a SOUL.md builder, a soul gallery (including character personas), preset quick-starts, an enforced heartbeat protocol, traffic-light status indicators, structured escalation, context protection and security inoculation. The pieces I'd reach for first are the enforced heartbeat (model + version display on every turn) and the never-forget / context-protection rules, since those map directly onto real reliability problems in long-running agents. It's dense and opinionated — plan on a real setup session rather than a drop-in.

Use case: Standing up a consistent, self-monitoring persona/ops layer for a long-running agent.

# Install the workspace contract before the personality This is a large operating system for an agent: personality documents, heartbeat protocol, escalation rules, context protection and memory directories. The striking part is not the personality catalogue — it is that the whole thing is grounded in a small set of state directories and a handful of standard shell tools. ## Create the state directories first The kit writes into a workspace with dedicated subdirectories for memory, archived memory and accumulated learnings. Create those paths before the first run, because a persona that cannot persist its memory degrades into a very confident amnesiac. ## Respect the declared tool requirements The skill declares dependencies on common shell utilities and treats the agent binary as optional. That ordering is deliberate: the management scripts work with plain tools, and only the parts that schedule work need the agent. When a setup step fails, check the declared binaries first rather than the persona configuration. ## The heartbeat is a protocol, not a schedule The heartbeat is specified as something that must display model and version information and report status through defined indicators. Follow the protocol literally at first. Improvising the heartbeat before the underlying state directories work turns every run into a slightly different behaviour, and a variable early step makes every later failure ambiguous. ## Keep escalation structured The kit defines an escalation path so the agent stops and hands over rather than improvising at the boundary of its authority. Wire that path to a real destination — a chat channel a human reads — before enabling autonomy features, because an escalation route nobody monitors is functionally the same as no escalation route.