qdrant-memory

name

qdrant-memory

description

Intelligent token optimization through Qdrant-powered semantic caching and long-term memory. Use for (1) Semantic Cache - avoid LLM calls entirely for semantically similar queries with 100% token savings, (2) Long-Term Memory - retrieve only relevant context chunks instead of full conversation history with 80-95% context reduction, (3) Hybrid Search - combine vector similarity with keyword filtering for technical queries, (4) Memory Management - store and retrieve conversation memories, decisions, and code patterns with metadata filtering. Triggers when needing to cache responses, remember past interactions, optimize context windows, or implement RAG patterns.

Qdrant Memory Skill

Token optimization engine using Qdrant vector database for semantic caching and intelligent memory retrieval.

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                      USER QUERY                              │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│  1. SEMANTIC CACHE CHECK (Cache Hit = 100% Token Savings)   │
│  ┌─────────────────┐    ┌─────────────────────────────────┐ │
│  │   Embed Query   │───▶│  Search Qdrant (similarity>0.9) │ │
│  └─────────────────┘    └─────────────────────────────────┘ │
│                                      │                       │
│                    ┌─────────────────┴──────────────────┐    │
│                    ▼                                    ▼    │
│            [CACHE HIT]                          [CACHE MISS] │
│            Return cached                        Continue to  │
│            response                             LLM          │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│  2. CONTEXT RETRIEVAL (RAG - 80-95% Context Reduction)      │
│  ┌─────────────────┐    ┌─────────────────────────────────┐ │
│  │  Identify Need  │───▶│  Retrieve Top-K Relevant Chunks │ │
│  └─────────────────┘    └─────────────────────────────────┘ │
│         Instead of 20K tokens ───▶ Only 500-1000 tokens     │
└─────────────────────────────────────────────────────────────┘

---

Prerequisites

Qdrant (Vector Database)

# Option 1: Docker (recommended)
docker run -d -p 6333:6333 -v qdrant_storage:/qdrant/storage qdrant/qdrant

# Option 2: Docker Compose (persistent)
# See references/complete_guide.md for docker-compose.yml

Embeddings Provider

Choose based on your needs:

Provider	Privacy	Cost	Speed	Setup
Ollama (recommended)	✅ Fully Local	Free	Fast (Metal)	brew install ollama
Bedrock (AWS/Kiro)	⚡ AWS Cloud	~$0.02/1M tokens	Fast	Uses AWS profile (no key)
OpenAI	❌ Cloud	~$0.02/1M tokens	Fast	API key required

Ollama Setup (M3 Mac Optimized)

# 1. Install Ollama (if not already installed)
brew install ollama

# 2. Start server (choose one option)
ollama serve              # Foreground (Ctrl+C to stop)
ollama serve &            # Background (current terminal)
nohup ollama serve &      # Background (survives terminal close)

# 3. Pull embedding model (768 dimensions, excellent quality)
ollama pull nomic-embed-text

# 4. Verify server is running
curl http://localhost:11434/api/tags

# 5. Test embedding generation
curl http://localhost:11434/api/embeddings -d '{"model":"nomic-embed-text","prompt":"hello"}'

Tip: To auto-start Ollama on login, add ollama serve & to your ~/.zshrc or use brew services start ollama.

Note: For Ollama, use --dimension 768 when creating collections.

Amazon Bedrock Setup (AWS/Kiro Subscription)

Uses your existing AWS credentials - no secrets stored in code.

# 1. Ensure AWS CLI is configured (uses ~/.aws/credentials)
aws configure  # Or set AWS_PROFILE for specific profile

# 2. Install boto3 if not present
pip install boto3

# 3. Set environment variables
export EMBEDDING_PROVIDER=bedrock
export AWS_REGION=eu-west-1  # Default region

# 4. Test authentication
python3 skills/qdrant-memory/scripts/embedding_utils.py

Models Available (cheapest first):

Model	Dimensions	Pricing
amazon.titan-embed-text-v2:0	1024	~$0.02/1M tokens
amazon.titan-embed-text-v1	1536	~$0.02/1M tokens
cohere.embed-english-v3	1024	~$0.10/1M tokens

Note: For Bedrock Titan V2, use --dimension 1024 when creating collections.

OpenAI Setup (Cloud)

export OPENAI_API_KEY="sk-..."

---

Quick Start

MCP Server Configuration

{
  "qdrant-mcp": {
    "command": "npx",
    "args": ["-y", "@qdrant/mcp-server-qdrant"],
    "env": {
      "QDRANT_URL": "http://localhost:6333",
      "QDRANT_API_KEY": "${QDRANT_API_KEY}",
      "COLLECTION_NAME": "agent_memory"
    }
  }
}

Initialize Memory Collection

Run scripts/init_collection.py to create the optimized collection:

# For Ollama (nomic-embed-text - 768 dimensions)
python3 scripts/init_collection.py --collection agent_memory --dimension 768

# For OpenAI (text-embedding-3-small - 1536 dimensions)
python3 scripts/init_collection.py --collection agent_memory --dimension 1536

---

Core Capabilities

1. Semantic Cache (Maximum Token Savings)

Purpose: Avoid LLM calls entirely for semantically similar queries.

Flow:

1. Embed incoming query
2. Search Qdrant for similar past queries (threshold > 0.9)
3. If match found → return cached response (100% token savings)
4. If no match → proceed to LLM, then cache result

Implementation:

# Cache check before LLM call
from scripts.semantic_cache import check_cache, store_response

# Check cache first
cached = check_cache(query, similarity_threshold=0.92)
if cached:
    return cached["response"]  # 100% token savings

# Generate response with LLM
response = llm.generate(query)

# Store for future cache hits
store_response(query, response, metadata={
    "type": "cache",
    "model": "gpt-4",
    "tokens_saved": len(response.split())
})

Collection Schema:

{
  "collection": "semantic_cache",
  "vectors": {
    "size": 1536,
    "distance": "Cosine"
  },
  "payload_schema": {
    "query": "keyword",
    "response": "text",
    "timestamp": "datetime",
    "model": "keyword",
    "token_count": "integer"
  }
}

2. Long-Term Memory (Context Optimization)

Purpose: Retrieve only relevant context instead of full conversation history.

Problem: 20,000 token conversation history → Expensive + Confuses model Solution: Query Qdrant → Return only top 3-5 relevant chunks (500-1000 tokens)

Memory Types:

Type	Payload Filter	Use Case
decision	type: "decision"	Past architectural/design decisions
code_pattern	type: "code"	Previously written code patterns
error_solution	type: "error"	How past errors were resolved
conversation	type: "conversation"	Key conversation points
technical	type: "technical"	Technical knowledge/docs

Implementation:

from scripts.memory_retrieval import retrieve_context

# Instead of passing 20K tokens of history:
relevant_chunks = retrieve_context(
    query="What did we decide about the database architecture?",
    filters={"type": "decision"},
    top_k=5,
    score_threshold=0.7
)

# Build optimized prompt with only relevant context
prompt = f"""
Relevant Context:
{relevant_chunks}

User Question: {user_query}
"""
# Now only ~1000 tokens instead of 20,000

3. Hybrid Search (Vector + Keyword)

Purpose: Combine semantic similarity with exact keyword matching for technical queries.

When to use: Error codes, variable names, specific identifiers

from scripts.hybrid_search import hybrid_query

results = hybrid_query(
    text_query="kubernetes deployment failed",
    keyword_filters={
        "error_code": "ImagePullBackOff",
        "namespace": "production"
    },
    fusion_weights={"text": 0.7, "keyword": 0.3}
)

---

MCP Tools Reference

Tool	Purpose
qdrant_store_memory	Store embeddings with metadata
qdrant_search_memory	Semantic search with filters
qdrant_delete_memory	Remove memories by ID or filter
qdrant_list_collections	View available collections
qdrant_get_collection_info	Collection stats and config

Store Memory

{
  "tool": "qdrant_store_memory",
  "arguments": {
    "content": "We decided to use PostgreSQL for user data due to ACID compliance requirements",
    "metadata": {
      "type": "decision",
      "project": "api-catalogue",
      "date": "2026-01-22",
      "tags": ["database", "architecture"]
    }
  }
}

Search Memory

{
  "tool": "qdrant_search_memory",
  "arguments": {
    "query": "database architecture decisions",
    "filter": {
      "must": [{ "key": "type", "match": { "value": "decision" } }]
    },
    "limit": 5,
    "score_threshold": 0.7
  }
}

---

Payload Filtering Patterns

Filter by Type

{
  "filter": {
    "must": [{ "key": "type", "match": { "value": "technical" } }]
  }
}

Filter by Project + Date Range

{
  "filter": {
    "must": [
      { "key": "project", "match": { "value": "api-catalogue" } },
      { "key": "timestamp", "range": { "gte": "2026-01-01" } }
    ]
  }
}

Exclude Certain Tags

{
  "filter": {
    "must_not": [
      { "key": "tags", "match": { "any": ["deprecated", "archived"] } }
    ]
  }
}

---

Collection Design Patterns

Single Collection (Simple)

agent_memory/
├── type: "cache" | "decision" | "code" | "error" | "conversation"
├── project: "<project_name>"
├── timestamp: "<ISO8601>"
└── content: "<text>"

Multi-Collection (Advanced)

Collection	Purpose	Retention
semantic_cache	Query-response cache	7 days
decisions	Architectural decisions	Permanent
code_patterns	Reusable code snippets	90 days
conversations	Key conversation points	30 days
errors	Error solutions	60 days

---

Token Savings Metrics

Track savings with metadata:

{
    "tokens_input_saved": 15000,
    "tokens_output_saved": 2000,
    "cost_saved_usd": 0.27,
    "cache_hit": True,
    "retrieval_latency_ms": 45
}

Expected Savings:

Scenario	Without Qdrant	With Qdrant	Savings
Repeated question	8K tokens	0 tokens	100%
Context retrieval	20K tokens	1K tokens	95%
Hybrid lookup	15K tokens	2K tokens	87%

---

Best Practices

Embedding Model Selection

Model	Dimensions	Speed	Quality	Use Case
text-embedding-3-small	1536	Fast	Good	General use
text-embedding-3-large	3072	Medium	Excellent	High accuracy
all-MiniLM-L6-v2	384	Fastest	Good	Local/private

Cache Invalidation

• Time-based: Expire cache entries after N days
• Manual: Clear cache when underlying data changes
• Version-based: Include model version in metadata

Memory Hygiene

1. Deduplicate: Check similarity before storing
2. Prune: Remove low-value memories periodically
3. Compress: Summarize long conversations before storing

---

References

• See references/complete_guide.md for full setup, testing, and troubleshooting
• See references/collection_schemas.md for complete schema definitions
• See references/embedding_models.md for model comparisons
• See references/advanced_patterns.md for RAG optimization patterns