MEOK Labs —
where we experiment in public.

What it is

The Dream State runs when the user is offline. The AI free-associates across its memory store — finding unexpected connections, generating candidate insights, and pre-loading creative contexts that might be useful next session. Think of it as the AI processing its day while you sleep. Results are logged as 'dream episodes' — the user can review them or dismiss them.

What we're testing

Whether unsupervised association across episodic memory produces retrievable insights with a higher novelty-relevance balance than standard RAG retrieval. We're using a bisociation scoring function (measuring conceptual distance × relational strength) to filter dream outputs before surfacing them.

What we've learned

The most useful dream outputs consistently bridge memories from different life domains — e.g., a work frustration and a childhood memory connected by a shared pattern. Pure recency-weighted RAG never surfaces these. Our bisociation filter now surfaces useful cross-domain insights at 16.7% hit rate — compared to 0% from standard RAG retrieval. Every dream cycle that runs makes the next one more precise.

What it is

Every sensitive action in MEOK — a response that might affect the user's emotional state, an irreversible task, a memory deletion — must pass a Byzantine fault-tolerant vote among the council. This experiment is the live production implementation: 33 specialist nodes across 6 tiers, weighted by domain expertise, running on a modified PBFT (Practical Byzantine Fault Tolerance) protocol adapted for LLM-agent contexts.

What we're testing

Whether PBFT consensus latency can be kept under 120ms at the 99th percentile in production LLM routing conditions — where node response times are orders of magnitude more variable than in classic distributed systems. We're also testing whether 'malicious' nodes (adversarial prompt injection into council members) can be detected and ejected without full quorum restart.

What we've learned

At 33 nodes with our current LLM-agent configuration, P99 consensus latency is 94ms — within target. Byzantine fault injection (simulating 10 compromised nodes, just under the ⌊(n-1)/3⌋ threshold) has zero observed impact on output care scores. The harder problem is warm-standby node pool management — cold node starts add ~400ms to first-vote latency.