Hebbian consolidation

The retrieval pipeline is only as good as the knowledge graph it operates on. Without an active maintenance process, the graph would accumulate dead weight from concepts creators explored once and abandoned, drift out of sync with actual platform activity, and fail to capture trends that emerge across the creator community. The nightly Hebbian consolidation cycle is the feedback loop that prevents this. It runs every night after creator session activity has settled, adjusting node weights based on actual usage and promoting cross-creator patterns to the shared universal graph.

Why Hebbian learning

The classic neuroscience principle: neurons that fire together, wire together. Applied to knowledge graphs:

Concepts that are consistently activated together (co-appear in tagged content, co-fire in retrieval sessions) gain stronger edge connections
Concepts that are rarely or never activated lose weight over time and eventually decay out of the active graph
The graph structure reflects actual creator behavior, not theoretical taxonomy design

This is the mechanism by which GenieHelper’s knowledge improves without manual curation. The system learns from what creators actually do.

The nightly schedule

00 UTC  →  node-decay.mjs         Hebbian decay on Directus user_nodes
30 UTC  →  taxonomy-hebbian.mjs   CO_OCCURS edge boost/decay on taxonomy_graph.json
45 UTC  →  taxonomy-reconcile.mjs Classify pending_review tags as active
00 UTC  →  fp_growth.mjs          Cross-user pattern mining (planned — B7-3)
30 UTC  →  promote.mjs            Write promoted patterns to Nodes/Transitional/ (planned)

Decay runs before pattern mining deliberately: stale signals are eliminated before the FP-Growth algorithm mines the cleaned graph. Amplifying stale nodes would produce false collective patterns.

Stage 1: node-decay.mjs

memory/consolidation/hebbian/node-decay.mjs applies Hebbian decay to every active node in the Directus user_nodes collection.

Decay formula

new_weight = weight × (1 − decay_rate) ^ days_since_last_reinforced

if new_weight < 0.05  →  status = 'decayed'

Each node stores its own decay_rate (default 0.05 — 5% per day). A node last reinforced 14 days ago with default decay rate retains:

1.0 × (1 - 0.05)^14 = 1.0 × 0.95^14 ≈ 0.488

After ~44 days without reinforcement, the node falls below the 0.05 threshold and is marked decayed.

Reinforcement

When a creator accepts generated content, the Action Runner calls node-decay.mjs reinforce <node_id>:

new_weight = min(1.0, weight + 0.1)
status = 'confirmed'
last_reinforced_at = now

A reinforcement event adds 0.1 to the weight (capped at 1.0) and resets the decay clock.

# Preview decay pass without writing (safe to run anytime)
node memory/consolidation/hebbian/node-decay.mjs --dry-run

# Reinforce a specific node after content acceptance
node memory/consolidation/hebbian/node-decay.mjs reinforce <node_id>

Stage 2: taxonomy-hebbian.mjs

memory/consolidation/hebbian/taxonomy-hebbian.mjs operates on the system-wide taxonomy graph (Nodes/Universe/taxonomy_graph.json). It boosts CO_OCCURS edge weights between tags that co-appeared in creator content during the previous cycle window, then applies global decay to all edges:

# Standard nightly run
node memory/consolidation/hebbian/taxonomy-hebbian.mjs

# Custom parameters
node memory/consolidation/hebbian/taxonomy-hebbian.mjs \
  --hours 48 --delta 2 --decay 0.99 --dry-run

--hours sets the lookback window for co-occurrence detection. --delta is the edge weight boost per co-occurrence. --decay is the global decay multiplier applied to all edges. Changes are written atomically to Nodes/Universe/taxonomy_graph.json to prevent partial writes during a crash.

Stage 3: cross-user FP-Growth pattern mining (planned)

Cross-user pattern mining is designed and structurally ready but not yet implemented (B7-3). The directory memory/consolidation/cross_user/ exists with its README. No code exists yet.

Once implemented, fp_growth.mjs will scan all confirmed user_nodes records across every creator and identify tag/concept combinations that co-occur above a configurable support threshold. FP-Growth parameters (design):

Parameter	Default	Description
`support`	≥ 0.15	Minimum fraction of creators where the pattern must appear
`confidence`	≥ 0.70	Minimum conditional probability for the association
`min_creator_count`	≥ 3	Minimum distinct creators where the pattern must appear before promotion

When a pattern clears all thresholds, promote.mjs writes it to Nodes/Transitional/ as a candidate for promotion to Nodes/Universe/.

Privacy design

Pattern mining aggregates across creators but never exposes individual attribution. Promoted patterns carry no creator IDs in the graph. Audit logs store pattern fingerprints only.

The node tier system

Knowledge nodes flow through three tiers, each with distinct semantics:

Nodes/Universe/

The canonical 3,205-node taxonomy graph. System-wide, shared across all creators. Written by taxonomy-hebbian.mjs and promote.mjs. Read by synaptic propagation at retrieval time.

Nodes/User/

Per-creator subgraphs. Each creator’s node activations, weights, and decay states. Read source for FP-Growth pattern mining. Written by node-decay.mjs via Directus user_nodes.

Nodes/Transitional/

Promotion staging area. Patterns that have cleared FP-Growth thresholds sit here awaiting human or automated review before moving to Nodes/Universe/. Prevents unreviewed patterns from polluting the universal graph.

agent-skills/ → DuckDB

Procedural memory. New skills drafted from observed agent behavior are ingested into agent_memory.duckdb (191 skills, 252 nodes, 12,880+ edges) and become retrievable on future sessions.

Nodes/User/{creator-uuid}/
    │
    │  nightly: FP-Growth mines confirmed nodes across all users
    ▼
Nodes/Transitional/
    │
    │  review (human or automated)
    ▼
Nodes/Universe/           ← 3,205-node canonical graph
    │
    │  taxonomy-hebbian.mjs applies nightly boost/decay
    ▼
Nodes/Universe/           ← updated graph, ready for retrieval

Automatic skill creation (in progress)

As the agent handles new request types, new skills are drafted to agent-skills/. Once ingested:

python3 memory/core/ingest_skills.py --reset
python3 memory/core/generate_hierarchical_goosehints.py

The new skills become nodes in DuckDB and are retrievable via JIT hydration on future sessions. The system accretes procedural memory — it gets better at tasks it has handled before.

Cron registration

# Add to crontab (crontab -e)
0 2 * * *  cd /var/www/vhosts/geniehelper.com/agentx && \
  node memory/consolidation/hebbian/node-decay.mjs \
  >> setup/logs/node-decay.log 2>&1

0 3 * * *  cd /var/www/vhosts/geniehelper.com/agentx && \
  node memory/consolidation/hebbian/taxonomy-hebbian.mjs \
  >> setup/logs/taxonomy-hebbian.log 2>&1

45 2 * * * cd /var/www/vhosts/geniehelper.com/agentx && \
  DIRECTUS_ADMIN_TOKEN=xxx \
  node memory/consolidation/taxonomy-reconcile.mjs \
  >> setup/logs/taxonomy-reconcile.log 2>&1

Logs are retained at 3,000 lines each. PM2 can also manage these via cron_restart in ecosystem.config.js.

These scripts have not yet been registered in the OS crontab or PM2 ecosystem config. Manual registration is required. Verify with crontab -l before assuming they are running nightly.

Implementation files

memory/consolidation/
├── hebbian/
│   ├── node-decay.mjs         ← per-creator Hebbian decay + reinforcement
│   ├── taxonomy-hebbian.mjs   ← system-wide CO_OCCURS edge boost/decay
│   └── README.md
├── cross_user/
│   └── README.md              ← FP-Growth design (no code yet)
└── taxonomy-reconcile.mjs     ← classifies pending_review tags

Key operations

Script	Operation	Target
`node-decay.mjs`	Decay: `weight × (1-rate)^days`	Directus `user_nodes`
`node-decay.mjs`	Reinforce: `weight + 0.1`, reset clock	Directus `user_nodes`
`taxonomy-hebbian.mjs`	Boost CO_OCCURS edges from recent co-occurrences	`Nodes/Universe/taxonomy_graph.json`
`taxonomy-hebbian.mjs`	Global edge decay pass	`Nodes/Universe/taxonomy_graph.json`
`taxonomy-reconcile.mjs`	Promote `pending_review` → `active`	Directus `taxonomy_mapping`
`fp_growth.mjs` (planned)	Mine frequent patterns across all `user_nodes`	Directus `user_nodes` (read)
`promote.mjs` (planned)	Write promoted patterns	`Nodes/Transitional/`

AI System

Memory & Retrieval

Taxonomy

Hebbian consolidation

Why Hebbian learning

The nightly schedule

Stage 1: node-decay.mjs

Decay formula

Reinforcement

Stage 2: taxonomy-hebbian.mjs

Stage 3: cross-user FP-Growth pattern mining (planned)

Privacy design

The node tier system

Nodes/Universe/

Nodes/User/

Nodes/Transitional/

agent-skills/ → DuckDB

Automatic skill creation (in progress)

Cron registration

Implementation files

Key operations

Build docs developers (and LLMs) love

AI System

Memory & Retrieval

Taxonomy

​Why Hebbian learning

​The nightly schedule

​Stage 1: node-decay.mjs

​Decay formula

​Reinforcement

​Stage 2: taxonomy-hebbian.mjs

​Stage 3: cross-user FP-Growth pattern mining (planned)

​Privacy design

​The node tier system

Nodes/Universe/

Nodes/User/

Nodes/Transitional/

agent-skills/ → DuckDB

​Automatic skill creation (in progress)

​Cron registration

​Implementation files

​Key operations

Build docs developers (and LLMs) love

Why Hebbian learning

The nightly schedule

Stage 1: node-decay.mjs

Decay formula

Reinforcement

Stage 2: taxonomy-hebbian.mjs

Stage 3: cross-user FP-Growth pattern mining (planned)

Privacy design

The node tier system

Automatic skill creation (in progress)

Cron registration

Implementation files

Key operations