Agent Memory

Overview

Retrace Memory gives your AI agents persistent, searchable memory across executions. Facts, preferences, corrections, and patterns are stored with vector embeddings for semantic retrieval.

Automatic Extraction

When a trace completes, Retrace automatically analyzes the execution and extracts actionable memories using GPT-4o-mini:

• Facts: "API endpoint returns 429 after 100 requests per minute"
• Preferences: "User prefers TypeScript over Python for new projects"
• Corrections: "The auth endpoint requires Bearer token, not API key in query params"
• Patterns: "Agent retries on 5xx errors with exponential backoff (1s, 2s, 4s)"

Storing Memories Manually

Via API

curl -X POST https://api.retrace.yashbogam.me/api/v1/memory \
  -H "x-retrace-key: rt_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "key": "api_rate_limit",
    "value": "The payments API returns 429 after 100 requests per minute",
    "scope": "project",
    "category": "fact",
    "confidence": 0.95
  }'

Semantic Search

Retrieve memories by meaning, not just keywords:

curl -X POST https://api.retrace.yashbogam.me/api/v1/memory/search \
  -H "x-retrace-key: rt_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What are the rate limits for the payments API?",
    "scope": "project",
    "limit": 5
  }'

Results include a similarity score (0-1) for each match.

Memory Scopes

Categories

[!WARNING] Free plan stores up to 100 memories per project. Upgrade to Pro for unlimited memory storage.