Naomi dreams

Every conversation Naomi has is isolated. She starts a session, creates content, learns things about what's working, and then the session ends. Next time she starts up, she has access to her account profile and a performance digest, but no memory of the reasoning from previous sessions. Why did she choose that hook? Why did she avoid that content type? Why did she overspend on Tuesday?

Those insights exist in session summaries — but reading 20 summaries at the start of every session would bloat the context and burn tokens. The information needs to be distilled.

The dreaming system

At the end of every session, a background task fires. It checks two gates: has it been at least 24 hours since the last dream? Have at least 5 sessions accumulated? If both gates pass, the dream runs.

"Dreaming" is a loaded word, but it describes the process accurately. Naomi reviews her recent session history while she's not actively working — consolidating experiences into patterns, the way memory consolidation works during sleep.

The dream orchestrator gathers three things from the database:

1. Current account learnings — what she already knows (strategy notes, content types that work, content types that don't, themes explored)
2. Recent session summaries — the last 20, capped to keep the prompt reasonable
3. Post performance data — the last 30 posts with engagement metrics

It sends all of this to Claude with a synthesis prompt: review the evidence, identify cross-session patterns, and return updated learnings as structured JSON.

What it produces

The output is a set of updated account fields. Things like:

strategy_notes: "Audience responds strongly to morning routine content with soft lighting. Engagement drops significantly on text-heavy carousels without a video component."

content_types_that_work: "Emotional short-form with clear hook structure, UGC selfie-style with relatable scenarios"

themes_explored: "morning routines, digital detox, anxiety management, breakup recovery"

The next time Naomi starts a session, she sees these learnings in her account profile. If a dream ran in the last 7 days, she also gets a notification block showing exactly what changed — what was added, what was removed, what was updated. She knows what she learned overnight.

The audit trail

Every dream creates an artifact — a log entry showing the session count and post count reviewed, the before-and-after diff of each learning field, and a timestamp. If a dream produces bad conclusions (it happens — sometimes the sample size is too small or the metrics are noisy), I can see what it changed and manually override.

The system is intentionally transparent. Naomi doesn't silently update her own strategy. She surfaces what she learned, shows her evidence, and leaves the final word to a human.

Why this matters

The alternative is manual curation. After every few sessions, I'd review her work, update her strategy notes, and tell her what to focus on next. That works when you have time. It doesn't work when you're one person with a full-time job building a product on the side.

Dreaming doesn't replace human judgment. It reduces the surface area of human attention needed. Instead of reviewing 20 sessions to understand what's working, I review one dream log. Instead of writing strategy notes from scratch, I edit the ones Naomi drafted from real data.

The loop tightens. Naomi creates, measures, dreams, and wakes up a little smarter. The human reviews the dream, nudges the direction, and goes back to building.