Guide

How to add persistent memory to Claude apps

Claude is powerful, but your application still needs durable user context. Mnexium adds memory, chat history, profiles, records, and agent state around Claude calls without forcing a custom memory stack.

1. Send Claude-compatible requests

Use Mnexium as the memory-aware API layer and include your Anthropic key in the request headers.

curl -X POST "https://mnexium.com/api/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -H "x-mnexium-key: $MNX_KEY" \
  -H "x-anthropic-key: $ANTHROPIC_API_KEY" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "messages": [
      { "role": "user", "content": "My name is Alex and I prefer concise answers." }
    ],
    "mnx": {
      "subject_id": "user_123",
      "chat_id": "chat_abc",
      "learn": true,
      "recall": true,
      "history": true
    }
  }'

2. Use stable IDs

The subject_id should map to the user, customer, account, or entity Claude should remember. The chat_id should map to the conversation thread.

  • Use the same subject_id when the same user returns.
  • Use the same chat_id when continuing a conversation.
  • Set learn to capture durable memories from the exchange.
  • Set recall to retrieve relevant memories before Claude responds.

3. Add structured context when needed

Memory is useful for preferences and facts. Records and profiles are better for structured application data such as accounts, orders, preferences, saved objects, and workflow state.

{
  "mnx": {
    "subject_id": "user_123",
    "recall": true,
    "records": {
      "recall": true,
      "learn": "auto",
      "tables": ["account", "project"]
    }
  }
}

Common mistakes

  • Using a new random subject_id on every request, which prevents continuity.
  • Saving everything as raw vector text when some context belongs in profiles or records.
  • Forgetting to enable recall on requests that should use prior memory.
  • Mixing provider keys and Mnexium keys in the wrong headers.

Next steps

Read the docsClaude memory deep diveAI memory API use case