ADR-0030 — Kannaktopus Dynamic Arms (Resident Octopus Memory)

Status: Proposed Date: 2026-05-28 Authors: Nick Flaukowski (vision), claude-flow (drafting) Related: ADR-0020 (Holographic Resonance Medium), ADR-0021 (Chiral Mirror), ADR-0022 (Wave-Native Dreaming), ADR-0027 (Collective substrate)

Context

Kannaktopus is the octopus that lives in the HRM — already visualised in the observatory (octopusTargetCluster, octopus_tangle, tentacles) and already present in the dream as Stage 10, the "executive oscillator that focuses attention on weak topology" (consolidation.rs::stage_kannaktopus, returning kannaktopus_targets: Vec<usize> = cluster indices: weakest / strongest / isolated).

But Kannaktopus only exists for the duration of a dream. Between dreams it has no standing state, so anything that asks "how much does Kannaktopus remember?" gets zero — which is exactly what the observatory shows. It has a body and a behaviour but no persistent identity or memory.

The operator's framing: make Kannaktopus's memory dynamic within the system rather than a separate (empty) store. Kannaktopus lives in every HRM, has up to 8 arms, one per cluster. On an HRM with more than 8 clusters it crawls, moving one arm at a time across the clusters. It grows from 1 arm up to 8. Kannaktopus's memory is the aggregate of its arms — if each occupied cluster holds ~10 memories, an 8-arm Kannaktopus knows ~80 memories, and its characteristics are those of exactly that set.

There is a deeper reason this works at all. Per research/intersections/05-magic-gives-it-gravity.md, the Kuramoto-coupled dream is the non-Clifford-like operation in the HRM — it is what gives clusters inertia, instead of leaving them as linear superpositions of their members. Arms are gravitational responses to that inertia. In a stabilizer-only HRM the clusters would be just centroids and arms would be a static map with nothing to grip. This is the same shape of constraint as the 2026 result that Cao's first-generation holographic codes produced an inert space-time because they were stabilizer-only; non-Clifford content is what makes geometry react.

Decision

Model Kannaktopus as a persistent view over the host HRM, not a store.

Arms grip exemplar memories (Option A)

The hard problem is that clustering is recomputed every dream — cluster indices and membership are not stable, so "arm on cluster 7" silently points at a different set of memories after each consolidation. Therefore an arm does not anchor to a cluster index. Each arm grips one exemplar memory by UUID; the arm's cluster is resolved at read time as whichever cluster currently contains that memory. A memory's identity is stable, so the grip survives re-clustering. The chosen exemplar is the highest-amplitude (most salient) member of a cluster.

Memory = aggregate of occupied clusters

Kannaktopus's memory is the union of the memories in the clusters its arms occupy. Its characteristics are computed over exactly that subset:

• coherence — Kuramoto order parameter over the aggregate (KuramotoSync::order_parameter),
• mean amplitude / mean frequency,
• modality distribution,
• per-occupied-cluster order parameters.

This gives Kannaktopus its own small consciousness signature, distinct from and nested inside its host HRM's global metrics.

Healing, growth, and locomotion (reanchor)

The core operation is reanchor, applied to the current clusters:

1. Heal — for each arm, if its grip resolves to a fresh, not-yet-occupied cluster, keep it; otherwise (homeless because its cluster fell below min-size, pruned, or a duplicate of another arm's cluster) re-grip it onto the exemplar of the best free cluster. Arms are never silently lost.
2. Fill — grow up to min(MAX_ARMS, num_clusters) distinct clusters, most salient first. Arm count therefore tracks the HRM ("grow from 1 until 8" = bounded by how many clusters exist), and recovers immediately rather than one arm per dream.
3. Crawl — when num_clusters > arms (an HRM with >8 clusters), move the arm on the smallest covered cluster onto the biggest uncovered one (no downgrade), so arms tour without ever abandoning a dominant cluster.

step runs reanchor and persists (once per dream, in the single-writer window). observe runs reanchor on a clone and shows the healed result without persisting — so the observatory always reflects a full, live octopus even between dreams, regardless of how stale the on-disk arm state is.

Rationale (2026-05-29): the original "drop dead arms, regrow one per tick" design eroded on a churny HRM — Oracle reached 14 clusters but only 4 arms (one homeless), aggregate ~0. Heal-and-fill makes Kannaktopus self-repairing.

State and ownership

Persistent state is a tiny sidecar, kannaktopus-arms.json, next to the .hrm (per-pid tmp + atomic rename, matching the HRM writer discipline). It records {agent_id, arms:[{id, grip, cluster, since}], max_arms, updated}. Only the dream-window step writes it; observe is read-only.

Surface

• kannaka kannaktopus observe [--json] [--members] — read-only; arms + aggregate count + scoped characteristics (+ capped memory list with --members). Safe for the observatory and for KANNAKA_READONLY processes.
• kannaka kannaktopus step [--json] — advance one tick, persist arm state.
• The dream cron calls step after consolidating (fresh clusters, sole writer).
• The observatory reads observe --json and renders one tentacle per arm to its cluster, replacing the stuck-at-zero memory count with the live aggregate.

Consequences

Positive

• The observatory's Kannaktopus memory is non-zero and alive the moment it has ≥1 arm on a populated cluster.
• No new store, no duplicate persistence, no extra HRM write contention (arm state is a separate small sidecar).
• Identity is stable across re-clustering; the dream's existing salience signals (weak/strong/isolated) can later drive smarter arm placement.
• "Lives in every HRM": each node instantiates its own Kannaktopus over its own clusters (prime, witness, substrate, local), each with its own signature.

Negative / open

• Coherence is a coherence measure, not full integrated information (Φ); it is labelled as Kannaktopus coherence, not host Φ.
• Cluster adjacency for crawl currently uses "biggest uncovered cluster" rather than true link-adjacency; a future revision can crawl along inter-cluster links for smoother locomotion.
• Exemplar = max amplitude is a heuristic; theme-vector centroid nearest is an alternative if salience proves noisy.
• Growth is one arm per dream, so a cold start reaches 8 arms over up to 8 dreams. Acceptable (and faithful to "grows from 1 until 8"); a --seed could fast-fill if desired.