[Feature] Relationship graph — lightweight edges between memories

[rahilp]

Why
Memories don't exist in isolation. "We decided to use Vectorize" is connected to "Vectorize has a 5MB metadata limit" which is connected to "we moved content storage to D1." Right now those connections only exist implicitly via semantic similarity. Making them explicit unlocks causality queries, dependency traversal, and timeline reconstruction.

What

• Edge types: Project → Decision, Decision → Outcome, Person → Project, Task → Status, Idea → Related Idea
• Edge creation: manual (user explicitly links two entries) and automatic (AI infers relationships at store time)
• Graph traversal in recall: when recalling entry X, optionally fetch all entries connected to X
  Implementation notes
• New D1 table: edges (id, source_id, target_id, relationship_type, created_at)
• Keep it lightweight — this is not a graph database, just a join table
• Graph traversal: simple recursive D1 query up to depth 2-3
• Web UI: show linked memories as "Related" section on memory detail view

[rahilp]

Merged from #17: Multi-hop recall

Why
Current recall is single-hop: query → top K matches. But many useful questions require following chains of reasoning. "What led to the decision to remove the search UI?" requires: find the search UI removal decision → find what informed it → find the context behind that. Multi-hop recall automates this chain.

What

• Query → retrieve top K → for each result, retrieve top K related → rerank the combined set globally
• Useful for: reconstructing rationale, tracing decision chains, understanding project history

Implementation notes

• Depends on relationship graph being in place for best results
• Can be approximated without graph: do a second vector query using the content of the top result as the new query
• Add hops parameter to recall tool (default 1, max 3)
• Higher hop counts increase latency significantly — cache aggressively

---

Merged from #17 — multi-hop traversal is a direct extension of the relationship graph and ships as part of the same feature.