Skip to content

How Context Works

SecondBrain does not treat "context" as just prompt stuffing.

Mental Model

There are three layers:

  1. source material
  2. promoted knowledge and memory
  3. task-specific runtime context

Source Material

This is what you ingest:

  • notes
  • docs
  • meeting transcripts
  • example datasets

The repo then maintains more durable structures:

  • promoted facts and assets
  • decision records
  • memory records
  • context cards and registries

Runtime Context

When you ask a question or run a task, the runtime assembles a smaller bundle:

  • the best matching evidence
  • prior decisions
  • relevant memory
  • policy and routing hints

That bounded bundle is what the planner or chat loop sees.

Why This Is Different From Plain Chat

  • evidence is retrieved and ranked, not pasted blindly
  • memory and decisions can influence the turn
  • policy and allowed actions can travel with the request
  • the final answer can be traced back to local state and source material