Skip to main content
The Context Engine is the picker. It looks at everything Memory Crystal knows and decides which pieces are most useful right now.

​
What this means in practice

Memory Crystal does not just dump all memory back into context. It uses ranking, filtering, and scope rules so the assistant sees the most relevant information first. That means the Context Engine cares about things like:
  • recency
  • similarity to the current question
  • memory strength
  • access history
  • graph connections
  • visibility or channel scope

​
How it actually works

The strongest public-facing grounding for this concept is in:
  • README.md
  • convex/crystal/recall.ts
  • convex/crystal/recallRanking.ts
  • convex/crystal/messages.ts
  • convex/crystal/mcp.ts
These files together show how Memory Crystal blends recall logic, ranking, message search, and tool/API surfaces.

​
Commands / examples

Useful adjacent docs:

​
Common mistakes

  • assuming vector search alone is the whole context system
  • assuming every relevant memory should always be injected
  • overlooking channel or scope isolation rules

​
Source of truth

Primary files behind this page:
  • README.md
  • convex/crystal/recall.ts
  • convex/crystal/recallRanking.ts
  • convex/crystal/messages.ts
  • convex/crystal/mcp.ts