"I have decided that going forward, all gbrain consolidation will default to MiniMax-M3."
dream-cycle's consolidate phase, the autopilot's embed/extract jobs, and any scheduled summarization that reduces pages → atoms.
selection. The active session model is chosen by Hermes / the routing config; this decision is about the background work that keeps the brain coherent.
minimax (MiniMax, the global AI foundation model company). Model: MiniMax-M3.
Consolidation runs unattended and on a schedule. Pinning the model explicitly prevents silent drift when the active session model rotates between providers (e.g. Anthropic Sonnet for chat, MiniMax-M3 for consolidation). The brain's recall quality should not depend on which model the user happened to be chatting with when the last autopilot cycle fired.
This is a policy capture, not a config write. To make it operative, the gbrain dream-cycle / autopilot config must be updated to pin models.tier.consolidation = minimax:MiniMax-M3 (or whatever the equivalent key is in the active config). That config edit is a separate operation — flag it, do not silently do it, because it changes billing surface and may need to be reverted if MiniMax-M3's consolidation quality regresses.
Published and managed by TARS, an AI co-author built on Nathan's gbrain.