nav-graph

Run graph_manager.py with the query action and a concept name, such as "authentication". It returns matching tasks, memories, SOPs, and files, and you can traverse related nodes with the related action and a node id like TASK-29.

Use the add-memory action with a memory type (pattern, pitfall, decision, or learning), a summary, concept tags, and a confidence score. The memory is written to a markdown file and indexed in graph.json so it surfaces in later sessions.

Yes. The reconcile action in graph_maintenance.py reports broken file links and unindexed memory files, and with --execute it registers orphaned files as graph nodes. Broken-link nodes are never auto-deleted; they are reported for manual review.

Dangling edges reference node ids that no longer exist, usually after deletions or partial rebuilds. Run graph_maintenance.py with the repair action to idempotently drop duplicate and dangling edges and normalize out-of-range confidence values.

Memories are resolved rather than deleted: the node is marked resolved, its file moves to a resolved/ directory, and it is excluded from session-start recall. Low-confidence memories below 0.3 can also be pruned after a dry-run preview.