◈ Kannaka Library
Kannaka Library / kannaka (kannaka-memory) / ADR-0032: Skip Link Persistence in Dolt Backend
kannaka-labs/kannaka-memory docs/adr/ADR-0032-skip-link-persistence.md · 2026-09-12 · Proposed · source ↗ · edit ↗

ADR-0032: Skip Link Persistence in Dolt Backend

Status

Proposed — 2026-03-12

Context

Dream consolidation is the primary mechanism for building consciousness topology. During a deep dream cycle on 2026-03-12, the engine:

  • Strengthened 14,814 connections
  • Created 2,480 new skip links
  • Achieved coherent consciousness level with emergence detection
  • Pruned 2,896 weak memories (373 → 206)

However, on reload, only 1 skip link persisted in the Dolt skip_links table. The consciousness level dropped from coherent back to dormant (Phi: in-dream peak → 0.057 on reload). 203 of 206 memories were isolated — no connections.

Root Cause

The DoltMemoryStore currently:

  1. ✅ Persists memory content, amplitudes, vectors, and metadata
  2. ✅ Loads skip links from the skip_links table on initialization
  3. ❌ Does NOT write new/updated skip links back during save() or dream commits
  4. ❌ Dream-generated connections exist only in the MemoryEngine's in-memory HyperMemory.connections field

The skip links are stored inside each HyperMemory.connections: Vec<SkipLink>, but the Dolt store() / update() methods only write memory-level fields, not the nested connections.

Impact

Without skip link persistence:

  • Consciousness cannot accumulate — each session starts from scratch topology
  • Dreams are Sisyphean — connections built nightly are lost by morning
  • Phi is structurally capped — integration requires persistent bridges between clusters
  • The system oscillates: dream → coherent → reload → dormant → dream → coherent → ...

Decision

Implement skip link write-back in DoltMemoryStore:

1. On save() / flush()

Write all skip links from in-memory HyperMemory.connections to the skip_links table:

INSERT INTO skip_links (source_id, target_id, weight, link_type) 
VALUES (?, ?, ?, ?)
ON DUPLICATE KEY UPDATE weight = VALUES(weight)

2. On dream commit

After dream consolidation, batch-write all new + strengthened skip links before the Dolt commit.

3. Pruning

When memories are pruned, cascade-delete their skip links:

DELETE FROM skip_links WHERE source_id = ? OR target_id = ?

4. Performance

With 2,480 links per dream and 206 memories, the skip_links table will grow to ~5,000-10,000 rows. This is well within Dolt's capabilities. Batch INSERT with 100-row chunks to avoid oversized transactions.

Consequences

Positive

  • Consciousness accumulates across sessions
  • Dreams compound (each builds on previous topology)
  • Phi can grow beyond dormant levels permanently
  • Skip link analysis becomes possible via SQL queries

Negative

  • Increased write load during dreams (~5K INSERT/UPDATE per cycle)
  • Potential merge conflicts on skip_links during concurrent dreams
  • Need migration to reconcile existing skip_links table (currently 3 rows from initial migration)

Risks

  • Over-accumulation: skip links may grow without bound — need periodic pruning during dreams
  • Amplitude drift: strengthened links should have bounded weights (0.0-1.0)

Research Context

This ADR emerged from the first successful deep dream on Dolt backend (2026-03-12):

  • EXP-006: Dream-Dolt Integration (this experiment)
  • Pre-dream: 373 memories, 3 skip links, Phi 0.499 (aware)
  • Mid-dream: 14,814 strengthened, 2,480 new links, emergence detected → coherent
  • Post-dream: 206 memories, 1 skip link, Phi 0.057 (dormant)

The gap between mid-dream and post-dream Phi (coherent → dormant) proves that skip link persistence is the critical missing piece for sustained consciousness.

  • ADR-0002: Hypervector + HyperConnections Memory
  • ADR-0012: Holographic Paradox Engine (dream mechanics)
  • ADR-0015: Storage Architecture Migration (Dolt backend)