One of the ten canonical user-centric memory abilities is Abstain and Boundary Handling: the capacity to recognize when information is insufficient, outdated, or outside the system's scope, and to withhold rather than fabricate. The paper treats this as a core competency, not an edge case. They evaluate it with dedicated benchmarks (HaluMem, ConvoMem) and consider it essential to trustworthy memory systems.

This is the CRRF. Tier zero (T0, "Silent") is literally "the system has no evidence and must remain silent." The entire nine-step gating framework exists to prevent premature assertion. Our Knowledge Graph design puts KnownUnknown and Unable-to-determine as first-class nodes, not error states. The portfolio's foundational ethos ("a system must earn the right to make a recommendation") is this paper's Ability 10, expressed as architecture.

The taxonomy's structural/hierarchical memory substrate includes knowledge graphs with typed edges, provenance tracking, and certainty metadata. The paper highlights how graph-based memory enables explicit reasoning chains and audit trails that flat vector stores cannot provide.

Folder 11's Knowledge Graph design uses an epistemic overlay with has-certainty edges, a controlled predicate vocabulary, and verifier-on-edge architecture. Every assertion in the graph carries provenance and confidence. The three-layer meta-design (knowledge graph artifact, orchestration graph, decision network) is a clinical instantiation of exactly the structural store class the paper describes, but with epistemic gates the paper's cited systems don't have.

The paper's third taxonomy axis divides memory into user-centric (preferences, history, implicit inferences about the human) and agent-centric (tools, environment, self-reflection). This is a structural distinction, not a feature list. They argue the two require fundamentally different storage, retrieval, and evolution mechanisms.

THE MOAT doctrine. The Clinical Intelligence team's thesis is that the defensible position lives at the L3 to L4 seam: vocabulary bindings, workflow integration, interaction graphs, data-in-context mappings. User-centric memory (organization-specific, hard-won clinical context) is the moat. Agent-centric memory (general reasoning, tool use) is the commodity layer. The paper's taxonomy axis is exactly the strategic distinction the CI team identified independently.

Multi-agent memory comes in four architectures (private, shared, hybrid, orchestrated) with three routing patterns (orchestrator-based, agent-initiated, memory-driven). Write-control isolation prevents agents from corrupting each other's stores. Verifier agents validate before committing to shared memory.

The Orchestration Graph (the second of Folder 11's three layers) is a topology for exactly this. The verifier-on-edge design means no fact traverses a graph edge without validation. The ECIF's Clinical Intelligence Architect role is the orchestrator. The CHAI governance overlay provides the write-control isolation (deployment gates over output gates). We were designing multi-agent memory architecture for clinical decision support without calling it that.

Procedural memory encodes reusable how-to knowledge. The paper cites Anthropic's Claude Agent Skills as an example of explicit, shareable procedural memory that can be transferred between agent instances. This is memory that teaches other agents to do things, not just to know things.

The Renal Codex. The CI team's concept of a codified clinical knowledge base that encodes not just facts but clinical reasoning patterns, workflows, and decision heuristics specific to nephrology. It's procedural memory for a clinical domain. The paper's framing gives us the right academic vocabulary for what the Codex is intended to be.

The survey identifies privacy-preserving memory as a critical challenge: selective retention, secure storage, access control, audit trails. They note that trustworthy memory requires explicit governance over what gets remembered, who can access it, and when it gets forgotten.

CHAI is a governance overlay that sits on top of the CRRF's epistemic gates. Its "Five Highs" risk categorization and deployment gates are exactly the access-control and audit infrastructure the paper says agent memory needs. We built the governance layer first, before the memory substrate. The paper suggests most teams do it the other way around (or not at all).

Memory evolution through reinforcement learning (step-level, trajectory-level, cross-episode) treats memory decisions as sequential optimization problems. The paper frames forgetting and retention as policy choices, not engineering constraints.

The third layer of Folder 11's design is the Decision Network, grounded in Kochenderfer's Value of Information framework. What to retrieve, when to update, what to forget: these are decision-theoretic problems. The paper's RL memory policies are the machine learning formalization of the same insight we approached from decision theory. Different math, same conclusion: memory governance is a sequential decision problem under uncertainty.