agentdb-causal-link

Link Causally

Add a typed edge between two memories so future queries can traverse cause→effect, supersedes, depends-on, etc.

When to use

• User says "X caused Y", "this depends on that", "this supersedes the old one"
• Post-incident analysis — recording what led to what
• Building a knowledge graph from a sequence of episodes
• ADR housekeeping — supersedes / amends / related-to between architecture decisions

API

agentdb_causal_edge(
  fromMemoryId:   <id>
  fromMemoryType: 'episode' | 'pattern' | 'skill' | 'adr'
  toMemoryId:     <id>
  toMemoryType:   <same set>
  relation:       'caused' | 'supersedes' | 'depends-on' | 'related-to' | <custom>
  similarity?:    0..1
  uplift?:        -1..1   // signed reward delta
  confidence?:    0..1
  metadata?:      { evidenceCount, sourceTimestamp, ... }
)

uplift is the secret sauce. Positive uplift = the causal link improved outcomes; negative = it hurt. The agentdb-investigator agent (this plugin) walks edges weighted by uplift × confidence to find root causes.

Cascade-aware delete

The agentdb 3.0.0-alpha.13 release added cascade: true semantics:

• agentdb_causal_edge_delete(edgeId) — single edge
• agentdb_causal_node_delete(nodeId, { cascade: true }) — node + all incident edges; returns deletedEdges count
• agentdb_edges_by_endpoints(from, to, label?) — bulk delete by tuple

Use cascade when removing an obsolete ADR or decommissioning an agent — leaving dangling edges in the graph poisons future traversals.

Don't

• Don't link memories without confidence < 0.5 unless you mark them as conjectures. Low-confidence edges accumulate and degrade agentdb_causal_explain quality.
• Don't use caused for correlation — reserve it for genuinely causal evidence (controlled change, repeated observation, expert confirmation).
• Don't forget the inverse direction. "A supersedes B" implies B is no longer authoritative; tag B's hierarchical entry as deprecated in the same operation.