Repository memory
Settings reference: behavior, defaults and safeguards for this page of desktop Settings.
Published unchanged from docs/settings/repository-memory.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1).
Technical reference for Settings / Repository memory. The app shows a short summary; this page keeps the full detail.
Repository memory (page help)
A persistent lexical-vector index and import/symbol graph, isolated by profile and actual repository identity. No hosted embedding provider or extra model key. Changed or ignored sources are marked stale, never silently treated as current.
Switches
Previously labeled "Enable explicit repository indexing and retrieval", "Allow bounded retrieved source snippets to enter model context" and "Allow reviewed Sky/Terra/Glacier knowledge in model context". Retrieved snippets are bounded in size.
Index scope
Source capture and freshness checks follow the selected session's Env root (Local, Sandbox or Remote). Without a session, project actions are explicitly Local. The derived text/vector cache stays on the control plane; selected snippets may enter the chosen model's context. Protected and ignored paths are excluded. Save before indexing; opening this page never scans or probes a host.
Remote warning
Pinned SSH repository memory, not a Local index or SendBox isolation. Inspect the bound index for its actual peer, root and build. Missing or changed hosts fail without fallback.
Index job status
A completed job means the complete captured index was published; inspect its binding and source digest. A stopped or failed job has settled; inspect its outcome and the last complete bound cache.
Coding-style knowledge
Save only non-sensitive project conventions. Knowledge outlives a session. Model observations require a proposal and your approval; they are never labeled user-authored facts.