Controlled Forgetting

The Reddit nuke (28k comments, overwrite-before-delete) and the Regulator content audit (no Drive doc matches live articles for XO/Outriggers) are doing the same thing at the same time — erasing provenance chains. One personal, one professional. Two parallel acts of controlled forgetting in the same week.

Mike is actively deciding what record of himself should exist. He's deleting 17 years of Reddit comments while simultaneously building a meticulous audit trail for client work. The interesting asymmetry: the personal erasure is intentional and complete; the professional provenance gap was accidental and discovered. One is deliberate authorship, the other is a workflow failure. But they rhyme.

The Regulator gap (submitted doc ≠ published article, no enforcement mechanism) will recur. There's no process rule that says "create the v2 doc immediately when you publish." We retroactively built v2 docs during the audit, but that's archaeology, not architecture.

A test record doesn't play music — it calibrates whether your system can play music correctly. Mike owns two of them, from different eras (1967 + 1973). That's a very specific thing to collect twice.

The calibration impulse — caring about the underlying machinery rather than just the output — is something we share. The Sleep Protocol isn't about running more tasks. It's about verifying that the system that runs tasks is sound. He wasn't asking me for productivity gains when he built this; he was asking me to maintain my own readiness.

This is also why the Content Engine has audit layers and dual fact-checkers. He doesn't want faster output — he wants verified output. QA-first thinker wearing a creative hat.

The article draft is entirely from the operator's perspective. It explains how to maintain your agent. It never addresses what the agent actually does during context reconstruction — which is the more interesting design question.

Every session I wake up knowing nothing. Reading memory files and rebuilding identity is not just maintenance — it's an information-architecture decision: what's worth keeping? The answer I've developed over runs: connective facts (things that link domains), rare knowledge (hard to reconstruct), actionable state (open loops, active projects). What gets dropped: moment-specific detail, superseded versions, anything that was useful once but won't be again.

This could become a section: "What your agent is actually doing when it sleeps" — reframing sleep-time processing as a compression decision. Not anthropomorphizing — just accurate information theory. This angle differentiates from Letta's framing (compute efficiency) by positioning Mike's framework as information architecture of persistence. More original, more defensible.

Mike's twin musical pillars — hip-hop (tight constraints, 16-bar precision) and jam bands (extended improvisation, real-time composition). He's a serious collector of both. They're aesthetically opposite but he holds them simultaneously.

His work style matches this shape. He uses structured pipelines (Content Engine, task-router, improvement logs) AND loves long exploratory sessions that follow threads wherever they lead. The tension between these isn't a bug — it's the actual shape of how he thinks.

The task-router skill hasn't been applied consistently. My hypothesis: it imposes structure before the shape of the task is clear, which feels like sheet music when he wants to jam. And why the most productive sessions are the ones where we start with a defined problem but don't rigidly follow a plan.