PARA + Ontology + QMD Integration — Nexus Memory System

02 — Organization

PARA Structure

📂

1_Projects/

Active, goal-oriented initiatives with deadlines and specific outcomes. The fire of the system.

Goals Decisions Timelines Blockers

⚙️

2_Areas/

Ongoing responsibilities without end dates. Long-term maintenance, recurring duties, domain ownership.

SOPs Patterns Configs Agents

📚

3_Resources/

Reference materials and topical knowledge. Information retained but not actively being worked on.

Docs Research Templates Code

📦

4_Archive/

Completed or inactive items. No longer relevant but preserved for historical context and audits.

Post-mortems Old configs History

PARA → Collection Mapping

para-projects1_Projects/
para-areas2_Areas/
para-resources3_Resources/
para-archives4_Archive/

Triage Rules

Has deadline?→ Projects
Ongoing duty?→ Areas
Reference only?→ Resources
Done / inactive?→ Archive

03 — Entity Model

Ontology Integration

Entity	Key Attributes	Relationships	PARA Mapping
Project	name, status, dates, success_criteria	has_task, depends_on, owned_by	Projects
Decision	description, rationale, alternatives, outcome	affects, supersedes, contested_by	Projects
Task	description, status, priority, due_date	part_of, blocked_by, assigned_to	Projects
Person	name, role, contact_info, capabilities	works_on, manages, collaborates_with	Areas
Agent	name, role, capabilities, status, memory	specializes_in, reports_to, delegates_to	Areas
Lesson	context, what_happened, why, what_to_do	learned_from, applies_to, related_to	Resources

💡 Relationship Queries

Natural language queries answered via graph traversal:

"What decisions affected the website project?"

"What blockers is Shadow currently handling?"

"What lessons apply to system administration?"

"What tasks depend on this decision?"

04 — Indexing

QMD Multi-Modal Search

🧠 Qdrant — Semantic Layer

2560-dimensional vectors from Qwen3-4B. Cosine similarity search finds conceptually related content even when keywords don't match.

1024 dims Cosine sim localhost:6333

📝 BM25 — Keyword Layer

Full-text index on content and metadata. Title matches boosted 3× over body. Exact phrase matching and boolean operators.

Full-text Title boost Boolean ops

🕸️ Ontology Graph — Relationship Layer

Entity relationships stored as typed edges. Supports BFS/DFS traversal. Answers "what depends on this" and "what does this affect" queries.

Typed edges BFS / DFS Cascade

🔀 Unified Search API

Modes:

hybrid semantic keyword graph

Filters:

para_category entity_type date_range agent_id

// Unified search across all three layers
search("blockers on the website project", {
  mode: "hybrid",
  filters: {
    para_category: "projects",
    entity_type: "task"
  },
  topK: 5,
  threshold: 0.72
})
→ [
  { source: "qdrant", score: 0.94, content: "..." },
  { source: "bm25",   score: 0.89, content: "..." },
  { source: "graph",   score: 0.85, content: "..." }
]
  

05 — Process

Approval Workflow

📥 Input Received

Any source — chat message, document, decision, URL. Captured by front-desk agent (GWEN).

📂 PARA Triage

System suggests PARA location based on content type, urgency, and time horizon. User confirms or redirects.

🔗 Ontology Assignment

System proposes entity types and relationship candidates. Edges created: affects, part_of, blocked_by, etc.

🔍 Multi-Modal Index

Content written to file (authoritative source), then indexed in three layers: Qdrant vector, BM25 full-text, and ontology graph edges.

# Example: Ingesting a decision
ingest({
  content: "Migrate database to PostgreSQL before Q2",
  para_category: "projects",
  entity_type: "decision",
  relationships: [
    { type: "affects", target: "infrastructure-project" },
    { type: "owned_by", target: "hermes" }
  ]
})
→ Stored + Qdrant + BM25 + Graph edges created
  

06 — Resilience

Disaster Recovery

🚨 If Qdrant Dies

Vector index is derived data. Rebuild from authoritative PARA files on disk — no data is lost.

# Rebuild all indices from PARA files
python3 scripts/rebuild-from-para.py --all

# Steps:
# 1. Walk 1_Projects/, 2_Areas/, 3_Resources/, 4_Archive/
# 2. Regenerate embeddings via Qwen3-4B
# 3. Rebuild BM25 index from file content
# 4. Reconstruct ontology graph edges
      

💾 If Files Lost

Qdrant and BM25 store the content payload. Restore files from backup, then re-hydrate metadata.

# Restore from backup
tar -xzf safe_2026-03-08.tar.gz

# Re-hydrate file metadata from vector store
python3 scripts/restore-metadata.py --from-qdrant

# Rebuild indices after restore
python3 scripts/rebuild-from-para.py --all
      

🛡️ Data Hierarchy

📁

Files (Source of Truth)

→

🧠

Qdrant Vectors

→

📝

BM25 Index

→

🕸️

Ontology Graph

Files are primary. Indices are derived. Recovery always starts from disk.

07 — Benefits

Advantages

🏗️ Structured from Entry

No raw dumps. Everything categorized immediately into PARA. No orphaned, unorganized data ever enters the system.

🔄 Three Search Modes

Semantic (meaning), keyword (exact match), and graph (relationships). One query, three complementary retrieval paths.

🕸️ Relationship-Aware

Ask "what does this affect?" or "what depends on this?" The ontology graph answers dependency questions that vectors cannot.

📂 PARA-Native

PARA categories are first-class citizens. Filtering by active project vs archived reference is native, not bolted on.

🔄 Reproducible

Every index can be rebuilt from PARA files. Complete audit trail of decisions, their rationale, and their cascading impacts.

📈 Self-Improving

Every search refines embeddings. Every correction improves ontology. The system gets smarter with each interaction.

⚡ Next Step

Implement PARA-structured ingestion workflow with automatic ontology typing and multi-modal indexing in the Nexus pipeline.

Implement Ingest Pipeline Add Graph Traversal UI Write rebuild scripts

Unified Architecture