Four forgetting modes plus budget-aware policy pruning.
Mode and target: $ARGUMENTS
act(action: "forget_node", node_id: "<id>", mode: "soft", reason: "Outdated reference")Node gets forgotten_at timestamp internally, excluded from retrieval, but structure/edges preserved. Reversible.
act(action: "forget_node", node_id: "<id>", mode: "hard")UtU-style constant-time unlink: severs all edges, deletes node. Irreversible.
act(action: "forget_node", node_id: "<id>", mode: "cascade", reason: "Removing stale test data")Also deletes orphaned dependents — nodes whose ONLY incoming support was from the deleted node.
act(action: "gdpr_erase", node_id: "<id>")Permanently removes node, all edges, and embedding data. Audit-logged per GDPR Article 17. No recovery possible.
Preview what would be forgotten:
act(action: "forget_policy", dry_run: true, max_nodes: 500)Execute the policy:
act(action: "forget_policy", max_nodes: 500)Priority scoring: confidence x recency x (1 + log(access_count + 1)) x (1 + connectivity). Lowest-priority nodes forgotten first via soft-forget.
| Scenario | Mode |
|---|---|
| Outdated but might be useful later | forget_node mode: soft |
| Definitely wrong, no dependents | forget_node mode: hard |
| Wrong + has orphaned dependents | forget_node mode: cascade |
| Contains PII / user data | gdpr_erase |
| Graph getting too large | forget_policy |
| Want to weaken, not remove | Use act(action: "belief_revise", operation: "contract") instead |
Don't GDPR-erase non-PII nodes — use soft/hard instead
Don't forget without checking dependents — use cascade if unsure
Always dry_run policy before executing
Don't forget nodes that are part of active goals — check goal linkage first