┌─────────────────────────────────────────────────────────┐
│                    CORTEX v3.2                          │
│                                                         │
│  ① WORKING MEMORY    temp_memories (Qdrant)             │
│     └─ quick_observe → instant write, no LLM           │
│                                                         │
│  ② SEMANTIC MEMORY   cortex_<project> (Qdrant)          │
│     └─ dense (all-MiniLM-L6-v2) + sparse (SPLADE)      │
│     └─ scored by qwen3 · linked · decay-weighted        │
│                                                         │
│  ③ EPISODIC MEMORY   cortex_episodes_<project> (Qdrant) │
│     └─ sessions with timestamped events, no LLM        │
└─────────────────────────────────────────────────────────┘

# 1. Qdrant (Docker)
docker run -d -p 6333:6333 --name qdrant qdrant/qdrant

# 2. Ollama + models
ollama pull qwen3:8b      # scoring, tagging, reranking, consolidation
# optional faster alternative:
# ollama pull qwen3:1.7b  # 3-5× faster on CPU, slightly lower accuracy

git clone git@github.com:alainrc2005/cortex_memory_mcp.git
cd cortex_memory_mcp
npm install
npm run build

QDRANT_URL=http://localhost:6333
FASTEMBED_CACHE_DIR=/absolute/path/to/cortex_memory_mcp/.fastembed_cache

# Optional — only needed if Qdrant has auth enabled
# QDRANT_API_KEY=your_key

# Optional — defaults to http://localhost:11434
# OLLAMA_URL=http://localhost:11434

{
  "mcpServers": {
    "langgraph-memory-mcp": {
      "command": "/absolute/path/to/cortex_memory_mcp/cortex-mcp.sh"
    }
  }
}

quick_observe(content)
      │
      ▼  (instant, no LLM, no embedding)
temp_memories  ◄──── working buffer (Qdrant, dummy vectors)
      │
      │  index_temp() — called at next session cold start
      ▼
  fastembed ONNX
  ├─ AllMiniLML6V2   → dense vector 384d
  └─ SpladePPEnV1    → sparse BM25 vector
      │
  qwen3 (Ollama)
  ├─ importance: 1-10
  ├─ type: DECISION | FACT | ERROR | PATTERN | PREFERENCE | CONTEXT
  └─ tags: [keyword, ...]
      │
  Contradiction detection (qwen3)
  └─ marks superseded memories if similarity > 0.88
      │
      ▼
  Qdrant upsert (dense + sparse vectors)
  └─ bidirectional links to related engrams

utility = alpha / (alpha + beta)
alpha += importance on access
beta  += 1 per day without access

stability     = log1p(accessCount) × (importance / 5)
retrievability = exp(-daysSinceAccess / stability)

src/
├── server.ts                    # MCP server — 20 tools, ~1500 LOC
├── bootstrap.ts                 # Qdrant collection init on startup
├── graph/
│   ├── observe/
│   │   ├── workflow.ts          # LangGraph pipeline: score→embed→link→persist
│   │   ├── state.ts             # Graph state types
│   │   └── nodes/
│   │       ├── score.ts         # qwen3: importance + type + tags
│   │       ├── embed.ts         # fastembed ONNX: dense + sparse vectors
│   │       ├── link.ts          # Bidirectional links in Qdrant
│   │       └── persist.ts       # Final upsert
│   └── consolidate/
│       └── nodes.ts             # LLM merge of duplicate engrams
├── services/
│   ├── qdrant.ts                # Qdrant client, collections, hybrid search
│   ├── fastembed.ts             # ONNX embeddings: dense (AllMiniLM) + sparse (SPLADE)
│   ├── ollama.ts                # qwen3: scoring, reranking, contradiction detection
│   ├── decay.ts                 # Dual decay engine: Bayesian + FSRS-inspired
│   └── episode.ts               # Episodic session management
└── types/
    ├── engrama.ts               # Engram TypeScript types
    └── episode.ts               # Episode/event types

npm test
# Covers all 20 tools with valid, invalid, and connectivity test cases

Session N (active)
    Agent detects storable fact
         ↓
    quick_observe()  ← instant, no CPU cost
         ↓
    Written to temp_memories

Session N+1 cold start
    get_context_for() + get_operator_profile()  ← parallel
         ↓
    Pending in temp_memories? → index_temp()
         ↓
    fastembed ONNX + qwen3 scoring applied
         ↓
    Promoted to cortex_<project> with full embeddings
         ↓
    Available for recall() and get_context_for()