Core concepts.

What belongs in memory

Empirical is for context that would need to be re-explained next week: coding preferences, architectural decisions, project constraints, team conventions. Skip transient values that won't improve a future session.

• ✓ Prefer ESM over CJS for new packages
• ✓ Auth uses client_credentials flow, not PKCE
• ✓ Monorepo: packages/shared must be built before packages/web
• ✗ Today's standup notes
• ✗ One-off calculation results or current prices

Workspaces

Memories belong to a workspace (default: personal). Switch workspaces to isolate project-specific context from personal preferences. Tools always read from the active workspace.

# List available workspaces
empirical workspace list

# Switch context
empirical workspace select

# MCP equivalent
set_current_workspace({ workspaceKey: "project-x" })

Relevance ranking

query_memories ranks results by semantic similarity, recency, and graph relationship weight (mass). High-mass memories — reinforced with notes, linked to related memories, or explicitly weighted — surface more readily than thin, isolated entries. You do not need to manage this manually; recording richer summaries and adding notes over time naturally increases signal.