{/* [LAYER: INFRASTRUCTURE] */}
An executive summary of LUMI's advisory workspace memory substrate.
Related: Knowledge Thesis · Knowledge Philosophy · Knowledge Whitepaper · README
Workspace Knowledge is LUMI's durable cognitive memory layer. Instead of relying on raw chat history or fragile vector embeddings to remember project details, LUMI builds a structured, typed model of the workspace after each completed task.
Task Execution ──► Task Finalization ──► learnFromFinalization()
│
▼
Next Task Context ◄── getKnowledgeHealth() ◄── Projections on Disk
The system tracks six key areas of project state:
- Subsystem Stability: Maps stable vs. high-churn volatile directories.
- Architectural Decisions: Logs ADR status and architectural progression.
- Stale Documentation: Points out documentation surfaces requiring review.
- Recurring Risk Areas: Identifies high-risk directories and code surfaces.
- Handoff Facts: Compiles key facts required for transition to the next agent.
Every fact contains a list of proof records indicating why the system believes it. It tracks files modified, the task run ID, the commit hash, and the exact timestamp. When state changes (e.g. a subsystem becomes stable), the fact is marked superseded and updated with a fresh provenance link.
Memory updates are completely advisory and non-blocking. If a file write fails (disk full) or a JSON file is corrupted, the system logs structured JSON warnings to diagnostics.jsonl, fallback renders the wiki page, and allows the task to successfully finish.
- Human-Readable Dashboard: View health status, recovery hints, and log entries directly at
.wiki/intelligence/workspace-intelligence.md. - Query APIs: Programmatic access via
WorkspaceIntelligenceReaderallows downstream tools and subagents to query health, ADR history, and directory volatility.