Description
Edge Lifecycle Automation keeps automatically generated relationships from becoming permanent clutter. It evaluates aging edges, lowers confidence on stale automatic connections, transitions sufficiently weakened edges to a decayed state, and prunes relationships that fall below the configured confidence policy. Maintenance stays company-scoped and uses the same lifecycle-aware service that handles normal graph mutation.
The feature distinguishes between machine-proposed edges and durable human or system assertions. Automatic discovery can be generous enough to make the graph useful, because lower-quality suggestions have a governed path out. Pruning preserves the audit trail rather than encouraging direct database deletion, and lifecycle changes remain observable to event consumers.
Selling Points
- Prevents old automatic suggestions from accumulating as unquestioned graph truth.
- Applies predictable confidence and age rules rather than ad hoc cleanup.
- Protects intentional user and system relationships from automatic decay behavior.
- Keeps maintenance company-scoped and authorization-aware.
- Emits lifecycle state through supported mutation paths for downstream consistency.
User Story
As a workspace administrator, I want discovered relationships to remain useful as our knowledge changes. I can rely on a configured cleanup policy to weaken and remove stale machine-generated links without asking people to audit every historical edge manually.
Extension Surface
The rating is configurable because company-level settings are loaded and updated through packages/knowledge-graph/src/config/company.ts, while packages/knowledge-graph/src/service.ts accepts bounded decay and prune options for maintenance runs. The weight and threshold rules live in packages/knowledge-graph/src/weights.ts. Configuration changes the policy but does not permit bypassing edge validation or deleting arbitrary graph rows.
Capabilities & Limits
Capabilities include time-window eligibility, stepwise confidence decay, status transition, threshold-based pruning, dry or bounded maintenance inputs, and event-aware deletion. The feature works naturally with confidence metadata produced by automatic discovery. Limits are deliberate: it does not decide that a manually authored fact is wrong, it does not recompute semantic similarity, and direct SQL removal remains unsupported because that would skip audit and edge-count behavior.
Implementation Map
packages/knowledge-graph/src/weights.tspackages/knowledge-graph/src/service.tspackages/knowledge-graph/src/config/company.tspackages/knowledge-graph/src/events.ts