Memory tree & Obsidian vault
- · Connectors stream into hierarchical summaries stored as SQLite on your machine.
- · Long-form content is chunked into ≤3k-token Markdown notes for LLM friendliness.
- · The Memory tab exposes the vault so you can edit or interlink pages like a Karpathy-style wiki.
1 · What Memory Tree stores
Marketing language calls it a Memory Tree: structured summaries built from all connected sources. Under the hood the critical detail is local persistence—your workflow knowledge remains on-device in SQLite while still allowing outbound model calls when you issue commands.
2 · Markdown chunks
Each integration's payload is canonicalized into Markdown segments capped around three thousand tokens so routing models receive consistent context windows. TokenJuice compresses HTML-heavy payloads before they ever hit the LLM.
3 · Obsidian workflow
The Memory tab includes an action like View vault in Obsidian, opening the on-disk wiki folder (documentation references a /wiki/ path—your build may show the resolved directory).
Manual links and edits you add inside Obsidian are fair game: the next ingest cycle can pick up those changes, so treat the vault as a shared scratchpad between you and the agent.
4 · Why this matters for prompts
Because context is pre-summarized, prompts like "What slipped through the cracks today?" scan across multiple connectors without you pasting raw email bodies. That only works after auto-fetch has populated the tree.
Continue to First prompts.