Do Agents Dream
of Electric Sleep?

Every day, Claude Code accumulates a little more entropy. Stale memory entries. Unused hooks. Bloated CLAUDE.md files. Repeated frustrations that never get addressed.

Today, this configuration rot is your problem. You're the one who notices the agent keeps making the same mistake. You're the one who cleans up.

But in biology, every organism has a solution to entropy. It's called sleep.

During sleep, your brain runs the most sophisticated maintenance cycle in nature:

• Memory consolidation — hippocampal replay strengthens important patterns
• Synaptic pruning — unused connections are weakened and removed
• Waste clearance — the glymphatic system activates 10x to clear metabolic debris
• Creative dreaming — REM sleep recombines ideas into novel solutions

Sleep is the price the brain pays for plasticity. Without it, the system degrades. With it, the system compounds.

Your immune system never attacks on a single signal. It requires co-stimulation — two independent signals confirming the threat before deploying a response.

Our sleep agent follows the same principle:

• First sighting — a pattern appears in one sleep cycle. Noted. Deferred.
• Second sighting — the same pattern appears again. Confirmed. Act on it.

This prevents hallucinated improvements from single-session noise.

Ant colonies coordinate through pheromone trails. Popular paths get reinforced. Abandoned paths evaporate. No central coordinator needed.

Every artifact in our system has a relevance score that decays each cycle:

Used artifacts get reinforced (+1.0). Unused artifacts fade. Below 0.2? Pruned.

This mirrors Tononi's synaptic homeostasis hypothesis: sleep exists to downscale synaptic strength, keeping only what matters.

Daytime — Braintrust CC plugin captures every session automatically. Online scorers analyze each trace in real-time.

Nighttime — The sleep cycle runs:

Each stage gets escalating capability. Haiku does cheap measurement. Sonnet does focused repair. Opus does creative synthesis.

Structured JSON output at every stage. Least-privilege tool access. Full traceability in Braintrust.

Same task. Before sleep and after. Measured in Braintrust.

The sleep agent created targeted memory entries, pruned stale context, and added rules that eliminated redundant tool calls. The agent woke up knowing things it had to rediscover before.

Built in one hackathon day, using Claude Code itself.

• sleep.sh — orchestrator chaining claude -p with JSON schemas
• 4 stage prompts — escalating model tiers and tool permissions
• query_braintrust.py — BTQL bridge between plugin and sleep agent
• upload_to_braintrust.py — 7 sessions, 23MB, 1,024 spans uploaded
• 3 online scorers — FrustrationDetector, TokenEfficiency, ToolUsageSummary
• 15,000+ lines of research across 300+ sources

300+ sources across neuroscience, immunology, ecology, and distributed systems:

• Tononi's synaptic homeostasis → trail evaporation
• Glymphatic waste clearance → orphaned file cleanup
• Hippocampal replay → cross-session aggregation
• REM dreaming → creative artifact synthesis
• Immune co-stimulation → two-signal activation
• Ant colony stigmergy → environmental traces as coordination
• Alan Kay's cathedral → simple components, powerful arrangement

• Scheduled execution — launchd triggers when idle for 30+ minutes. True healing while you sleep.
• Multi-project learning — same correction across 3 repos? Strong signal.
• Continuous cycles — repeat N1→REM like real sleep (4-5 cycles per night) with diminishing returns detection.
• Community patterns — 1,000 developers frustrated by the same thing? Anthropic should know.
• Voice briefing — "Good morning. While you slept, I pruned 3 stale memories and created a new skill. Token usage should drop 40%."