DMH-AI

API Pools

Assistant

Confidant

Swift — short, fast actions

Used for: anchor-pivot classify, runtime message localisation, web-search query planning, session naming. Cheapest model — latency dominates over quality.

Oracle — long, dense content

Used for: context compaction, task progress summaries, web-result synthesis, profile extraction + condensation. Strong general model.

Vision — image / video / OCR

Used for: image describer, video describer, PDF OCR. Must be image-capable.

Vector Embedding

MCP Catalog

Curated Services

Wiki Seeds

Token Saving

Conservative token saving — flush closed tasks from chat context. The model still sees the Task List and can fetch_task(N) for details on demand.

Connected accounts

Chat

Compact after messages

Keep recent messages

Max tool result chars

Trace LLM calls and tool results to file

Multimedia

Video analysis depth

Companion Memory

Condense after facts