Skip to content

Performance Tuning

Jump to bottom
Zhafron Kautsar edited this page Jan 11, 2026 · 1 revision

Performance Tuning

Optimization strategies for OpenCode Memory performance, resource usage, and cost efficiency.

Performance Metrics

Key metrics to monitor:

  • Search latency
  • Memory usage (RAM)
  • Database size
  • API costs (if using external services)
  • Auto-capture frequency

Search Performance

Reduce Search Latency

Lower Memory Limits:

{
  "maxMemories": 3,
  "maxProjectMemories": 5
}

Fewer results = faster search.

Increase Similarity Threshold:

{
  "similarityThreshold": 0.7
}

Higher threshold filters more results, reducing processing.

Use Smaller Embedding Model:

{
  "embeddingModel": "Xenova/all-MiniLM-L6-v2"
}

384 dimensions vs 768 = 2x faster.

Optimize Shard Size:

{
  "maxVectorsPerShard": 25000
}

Smaller shards = faster individual queries.

Benchmark Results

Search Time by Model (10,000 memories):

Model Dimensions Time
Xenova/all-MiniLM-L6-v2 384 15ms
Xenova/nomic-embed-text-v1 768 28ms
Xenova/bge-base-en-v1.5 768 30ms
Xenova/bge-large-en-v1.5 1024 45ms

Search Time by Database Size (768 dimensions):

Memories Time
1,000 8ms
10,000 28ms
50,000 85ms
100,000 160ms

Memory Usage (RAM)

Reduce RAM Consumption

Use Smaller Embedding Model:

{
  "embeddingModel": "Xenova/all-MiniLM-L6-v2"
}

RAM usage: 200MB vs 500MB+ for larger models.

Enable Auto-Cleanup:

{
  "autoCleanupEnabled": true,
  "autoCleanupRetentionDays": 14
}

Keeps database small.

Reduce Shard Size:

{
  "maxVectorsPerShard": 25000
}

Smaller shards use less memory per query.

Disable Web Server (if not needed):

{
  "webServerEnabled": false
}

Saves ~50MB RAM.

RAM Usage by Component

Component RAM Usage
Base plugin 50MB
Embedding model (small) 200MB
Embedding model (medium) 500MB
Embedding model (large) 2GB
Web server 50MB
Database cache 100MB

Database Size

Reduce Disk Usage

Run Cleanup Regularly:

curl -X POST http://127.0.0.1:4747/api/cleanup \
  -H "Content-Type: application/json" \
  -d '{"retentionDays":30,"dryRun":false}'

Run Deduplication:

curl -X POST http://127.0.0.1:4747/api/deduplicate \
  -H "Content-Type: application/json" \
  -d '{"similarityThreshold":0.9,"dryRun":false}'

Vacuum Database:

sqlite3 ~/.opencode-mem/data/memories_shard_0.db "VACUUM;"

Reclaims deleted space.

Use Smaller Dimensions:

{
  "embeddingModel": "Xenova/all-MiniLM-L6-v2",
  "embeddingDimensions": 384
}

384 dimensions = 50% smaller than 768.

Database Size Estimates

Per Memory (average):

  • Content: 200 bytes
  • Metadata: 100 bytes
  • Vector (384d): 1.5 KB
  • Vector (768d): 3 KB
  • Vector (1024d): 4 KB

Total Database Size:

Memories 384d 768d 1024d
1,000 2 MB 3 MB 4 MB
10,000 18 MB 32 MB 42 MB
50,000 88 MB 156 MB 208 MB
100,000 175 MB 310 MB 415 MB

API Cost Optimization

Reduce Embedding API Costs

Use Local Models:

{
  "embeddingModel": "Xenova/nomic-embed-text-v1"
}

Zero API costs.

Switch to Cheaper API Model:

{
  "embeddingModel": "text-embedding-3-small",
  "embeddingApiUrl": "https://api.openai.com/v1"
}

$0.02 per 1M tokens vs $0.13 for large model.

Reduce Auto-Capture Costs

Increase Token Threshold:

{
  "autoCaptureTokenThreshold": 20000
}

Halves capture frequency.

Use Cheaper Model:

{
  "memoryModel": "gpt-3.5-turbo"
}

90% cost reduction vs GPT-4.

Reduce Context Window:

{
  "autoCaptureContextWindow": 2
}

Fewer tokens per capture.

Limit Max Memories:

{
  "autoCaptureMaxMemories": 5
}

Shorter responses = lower cost.

Cost Estimates

Embedding Costs (OpenAI text-embedding-3-small):

Memories Cost
1,000 $0.002
10,000 $0.02
100,000 $0.20

Auto-Capture Costs (GPT-4, 10k threshold):

Daily Tokens Captures/Day Monthly Cost
50,000 5 $4.50
100,000 10 $9.00
200,000 20 $18.00

Auto-Capture Costs (GPT-3.5-Turbo, 10k threshold):

Daily Tokens Captures/Day Monthly Cost
50,000 5 $0.45
100,000 10 $0.90
200,000 20 $1.80

Configuration Profiles

Speed-Optimized

{
  "embeddingModel": "Xenova/all-MiniLM-L6-v2",
  "similarityThreshold": 0.7,
  "maxMemories": 3,
  "maxProjectMemories": 5,
  "maxVectorsPerShard": 25000
}

Characteristics:

  • Fast search (10-15ms)
  • Low RAM (250MB)
  • Good quality

Quality-Optimized

{
  "embeddingModel": "Xenova/bge-large-en-v1.5",
  "similarityThreshold": 0.5,
  "maxMemories": 10,
  "maxProjectMemories": 20,
  "maxVectorsPerShard": 50000
}

Characteristics:

  • Best search quality
  • Higher RAM (2GB+)
  • Slower search (40-50ms)

Balanced

{
  "embeddingModel": "Xenova/nomic-embed-text-v1",
  "similarityThreshold": 0.6,
  "maxMemories": 5,
  "maxProjectMemories": 10,
  "maxVectorsPerShard": 50000
}

Characteristics:

  • Good balance
  • Moderate RAM (500MB)
  • Moderate speed (25-30ms)

Low-Resource

{
  "embeddingModel": "Xenova/all-MiniLM-L6-v2",
  "similarityThreshold": 0.7,
  "maxMemories": 3,
  "maxProjectMemories": 5,
  "maxVectorsPerShard": 25000,
  "autoCleanupEnabled": true,
  "autoCleanupRetentionDays": 14,
  "deduplicationEnabled": true
}

Characteristics:

  • Minimal RAM (250MB)
  • Small database
  • Fast search
  • Automatic cleanup

Cost-Optimized

{
  "embeddingModel": "Xenova/nomic-embed-text-v1",
  "autoCaptureEnabled": true,
  "memoryModel": "gpt-3.5-turbo",
  "autoCaptureTokenThreshold": 20000,
  "autoCaptureContextWindow": 2,
  "autoCaptureMaxMemories": 5
}

Characteristics:

  • Zero embedding costs (local)
  • Low auto-capture costs
  • Good quality

Monitoring

Check Performance

Search Latency:

time curl -X POST http://127.0.0.1:4747/api/search \
  -H "Content-Type: application/json" \
  -d '{"query":"test"}'

Database Size:

du -sh ~/.opencode-mem/data/

Memory Usage:

ps aux | grep opencode

Statistics Dashboard

Access web interface statistics:

http://127.0.0.1:4747

Click "Statistics" tab for:

  • Memory counts
  • Database size
  • Search performance
  • Auto-capture stats

Benchmarking

Run Benchmark

Create test script:

#!/bin/bash
for i in {1..100}; do
  curl -s -X POST http://127.0.0.1:4747/api/search \
    -H "Content-Type: application/json" \
    -d '{"query":"test query"}' > /dev/null
done

Measure average time:

time ./benchmark.sh

Compare Configurations

  1. Benchmark with current config
  2. Change configuration
  3. Restart OpenCode
  4. Benchmark again
  5. Compare results

Best Practices

Regular Maintenance

Weekly:

  • Check database size
  • Review statistics

Monthly:

  • Run cleanup
  • Run deduplication
  • Vacuum database

Quarterly:

  • Review configuration
  • Benchmark performance
  • Consider model upgrade

Optimization Strategy

  1. Measure Current Performance
  2. Identify Bottleneck (search, RAM, disk, cost)
  3. Apply Targeted Optimization
  4. Measure Again
  5. Iterate

Avoid Over-Optimization

Don't optimize prematurely:

  • Default config works well for most users
  • Only optimize if experiencing issues
  • Balance performance vs quality vs cost

Next Steps

Clone this wiki locally