ADR-049: Self-Learning Memory with GNN & RuVector Integration

Status: Implemented Date: 2026-02-08 Authors: RuvNet, Claude Flow Team Supersedes: None Related: ADR-048 (Auto Memory Integration), ADR-006 (Unified Memory), ADR-009 (Hybrid Memory Backend)

Context

ADR-048 established the AutoMemoryBridge for bidirectional sync between Claude Code auto memory files and AgentDB. While this successfully bridges the two systems, it operates as a passive store — insights are recorded but the system does not learn from them. Three gaps exist:

1. No learning pipeline: The @claude-flow/neural package has a fully implemented NeuralLearningSystem with SONA, ReasoningBank (4-step RETRIEVE/JUDGE/DISTILL/CONSOLIDATE pipeline), and PatternLearner — but these are completely disconnected from the memory bridge.

2. No knowledge graph: MemoryEntry.references supports graph relationships between entries, but nothing constructs or queries a graph from them. Insights are flat lists without structural understanding.

3. No agent scoping: Claude Code supports 3-scope agent memory directories (project, local, user) for per-agent knowledge isolation, but the bridge only handles the project-level auto memory directory.

Decision

Extend the AutoMemoryBridge with three new modules:

1. LearningBridge (learning-bridge.ts)

Connects insights to the neural learning pipeline:

Insight Recorded → Begin Trajectory → Record Steps
Insight Accessed → Boost Confidence → Record Step
Consolidate      → Complete Trajectories → JUDGE/DISTILL/CONSOLIDATE
Time Passes      → Decay Confidences

• Optional dependency: @claude-flow/neural is loaded dynamically; when unavailable, all learning operations degrade to no-ops (confidence remains static).
• Confidence lifecycle: Entries gain confidence when accessed (+0.03 per access, capped at 1.0) and lose confidence over time (-0.005/hour, floored at 0.1).
• Consolidation: Triggered during session-end sync. Completes accumulated trajectories, runs the ReasoningBank pipeline, and updates entry metadata.

2. MemoryGraph (memory-graph.ts)

Builds an in-memory knowledge graph from entry references and similarity:

Build from Backend → Add Nodes → Add Reference Edges
                                → Add Similarity Edges (HNSW)
Compute PageRank   → Power Iteration (d=0.85, convergence=1e-6)
Detect Communities → Label Propagation
Rank with Graph    → alpha * vectorScore + (1-alpha) * normalizedPageRank

• Pure TypeScript: PageRank via power iteration and community detection via label propagation — no external graph libraries.
• Graph-aware curation: curateIndex() can prioritize high-PageRank entries for MEMORY.md, ensuring the most connected/influential insights appear first.
• Edge types: reference (from MemoryEntry.references), similar (from HNSW search), temporal, co-accessed, causal.

3. AgentMemoryScope (agent-memory-scope.ts)

Maps Claude Code's 3-scope agent memory system:

project: <gitRoot>/.claude/agent-memory/<agentName>/
local:   <gitRoot>/.claude/agent-memory-local/<agentName>/
user:    ~/.claude/agent-memory/<agentName>/

• Knowledge transfer: High-confidence insights (>0.8) can be transferred between agent scopes, enabling cross-agent learning.
• Path sanitization: Agent names are sanitized to prevent path traversal.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Claude Code Session                      │
│  ┌──────────────────┐  ┌──────────────────┐                │
│  │  Auto Memory      │  │  Agent Memory     │                │
│  │  (MEMORY.md)      │  │  (3-scope dirs)   │                │
│  └────────┬─────────┘  └────────┬─────────┘                │
│           │                      │                           │
│  ┌────────▼──────────────────────▼─────────┐                │
│  │           AutoMemoryBridge               │                │
│  │  ┌──────────────┐  ┌─────────────────┐  │                │
│  │  │ LearningBridge│  │  MemoryGraph    │  │                │
│  │  │  (optional    │  │  (PageRank +    │  │                │
│  │  │   neural)     │  │   communities)  │  │                │
│  │  └──────┬────────┘  └──────┬──────────┘  │                │
│  │         │                  │              │                │
│  │  ┌──────▼──────────────────▼──────────┐  │                │
│  │  │          AgentDB (HNSW)            │  │                │
│  │  │     150x-12,500x faster search     │  │                │
│  │  └────────────────────────────────────┘  │                │
│  └──────────────────────────────────────────┘                │
│           │                                                   │
│  ┌────────▼─────────────────┐                                │
│  │  @claude-flow/neural     │  (optional peer dependency)    │
│  │  - NeuralLearningSystem  │                                │
│  │  - SONA + ReasoningBank  │                                │
│  │  - PatternLearner        │                                │
│  └──────────────────────────┘                                │
└─────────────────────────────────────────────────────────────┘

Integration Points

AutoMemoryBridge modifications (+70 lines)

• config.learning?: LearningBridgeConfig — enables learning
• config.graph?: MemoryGraphConfig — enables graph
• recordInsight() → calls learningBridge.onInsightRecorded()
• syncToAutoMemory() → calls learningBridge.consolidate() first
• curateIndex() → uses graph.getTopNodes() for section ordering
• importFromAutoMemory() → builds graph from imported entries
• destroy() → calls learningBridge.destroy()

Hooks integration

• session-start → importFromAutoMemory(), build graph
• session-end → consolidate(), syncToAutoMemory(), curateIndex()
• post-task → recordInsight() for task learnings

Performance Targets

Testing Strategy

• TDD London School: All dependencies mocked (IMemoryBackend, NeuralLearningSystem)
• Graceful degradation: Tests verify no-op behavior when @claude-flow/neural unavailable
• Existing tests preserved: 73 AutoMemoryBridge tests must remain green
• Target: 100+ new tests across 3 modules

File Summary

Total: 219 tests passing in 385ms across 4 test suites

Implementation Status

Phase 1: LearningBridge -- COMPLETED

• LearningBridge class with neuralLoader injection (453 lines)
• Trajectory tracking (insight → trajectory mapping)
• Confidence boost on access + time-based decay
• Consolidation pipeline (JUDGE/DISTILL/CONSOLIDATE)
• Pattern search via ReasoningBank
• 56 tests passing (54ms)

Phase 2: MemoryGraph -- COMPLETED

• MemoryGraph class with PageRank + label propagation (392 lines)
• Graph construction from MemoryEntry.references
• Similarity edge auto-creation via HNSW search
• Graph-aware ranking (alpha-blended vector + PageRank)
• BFS neighbor traversal with depth control
• 60 tests passing (24ms)

Phase 3: AgentMemoryScope -- COMPLETED

• 3-scope path resolution with traversal protection (300 lines)
• createAgentBridge() factory for scoped bridges
• transferKnowledge() with confidence filtering + content-hash dedup
• listAgentScopes() for scope discovery
• 30 tests passing (22ms)

Phase 4: Integration -- COMPLETED

• AutoMemoryBridge config extended with learning and graph options
• Wired learningBridge into recordInsight, syncToAutoMemory, destroy
• Wired memoryGraph into importFromAutoMemory, curateIndex
• All exports added to index.ts and types.ts
• 73 existing AutoMemoryBridge tests still passing (no regressions)

Consequences

Positive

• Insights now trigger learning trajectories, improving curation over time
• PageRank identifies the most structurally important insights for MEMORY.md
• Agent-scoped memory enables per-agent knowledge isolation and transfer
• Fully optional: all new features degrade gracefully when dependencies unavailable

Negative

• Additional complexity in AutoMemoryBridge (~70 lines)
• Graph construction adds startup latency (~200ms for 1k entries)
• @claude-flow/neural becomes an optional peer dependency

Risks

• Neural system changes may require LearningBridge updates (mitigated by dynamic import + try/catch)
• Large graphs (>5k nodes) may impact memory usage (mitigated by maxNodes cap)