[PIR][WP18] Cross-source causal-graph memory fusion layer (MemFuse pattern, ADR-320)

Part of the RuV Perpetual Intelligence Runtime (PIR), Wave 2. Epic: #837. Plan: docs/research/perpetual-intelligence-runtime/07-wave2-program-plan.md · evidence: 06-wave2-evidence-review.md (both land on main with the feat/pir-wave2-adrs PR).

ADR mapping

Proposed ruvector ADR-320 — cross-source causal-graph fusion layer on the continuous-latent-state memory tier. Extends merged ruvector ADR-307 (three-level memory — specifically the continuous-latent-state tier, WP3 #839) and ruvector ADR-310 (causal-attribution gate, WP6 #850). Fuses atomic, source-tagged observations (AtomicObservation) from multiple agents/sensors into one causal graph with provenance back to the originating event — extending ADR-307's single-agent design to the Latent Communication Fabric's multi-agent setting.

Evidence

MemFuse (arXiv:2608.18704, Darwin-Agent/Mi-Memory) — "MemFuse: Multi-Source Memory Fusion from Fragmented Observations," submitted 2026-08-19. Grade A — qualitative claim confirmed. Artifact available now: github.com/Darwin-Agent/Mi-Memory/tree/master/MemFuse (verified real content; an earlier pass wrongly suspected no code existed) — adapt the released implementation rather than rebuilding from the paper; benchmark against the released MemFuseBench.

⚠️ Naming collision — SEVERE (the most important finding of the Wave-2 evidence review): github.com/memfuse/memfuse is an established, actively-forked, unrelated open-source LLM memory layer — a direct conceptual competitor under the same name. Never ship an npm package, crate, or module literally named memfuse. Always cite the paper as "MemFuse (arXiv:2608.18704, Darwin-Agent/Mi-Memory)", explicitly distinct from the pre-existing memfuse/memfuse project, in every ADR and code comment.

Scope

• Add the cross-source causal-graph fusion layer to the continuous-latent-state memory tier, adapting the released implementation.
• Surfaces: agentdb, crates/rvf (continuous-latent-state tier); coordinates with latentmesh-align-consuming code per the existing ADR-310 CI gate — this WP's fusion output feeds the same causal audit.

Candidate-mutation rule (applies to all of Wave 2)

Candidate mutation, not trusted prior art. Promotion requires this program's own research-gate paired-bootstrap-recomputed delta over the pre-WP baseline — never the paper's self-reported numbers.

Acceptance criteria

• Fusion layer implemented on the continuous-latent-state tier with per-observation source provenance (ADR-320)
• Internal MemFuseBench run shows a research-gate-measured delta over the pre-WP baseline
• No artifact named memfuse anywhere in the deliverable; citation discipline verified in review
• Fusion output flows through the ADR-310 causal-audit gate
• ADR-320 merged with repo-qualified citations per ADR-305 §4

Depends on: WP3 (#839) memory tiers, WP6 (#850) causal-audit gate. Phase W2-2 — starts after the W2-1 trio (WP15/16/17).

Read docs/research/perpetual-intelligence-runtime/07-wave2-program-plan.md and 06-wave2-evidence-review.md first, then inspect the continuous-latent-state surfaces in agentdb and crates/rvf. Review ADR-307, ADR-310, and the released MemFuse artifact while confirming the WP3 and WP6 dependencies. Done means the fusion layer, provenance, research-gate benchmark, causal-audit flow, naming checks, and ADR-320 acceptance criteria are complete.