Paris MongoDB User Group

MongoDB as your LLM agent's
audit trail

and why it makes debugging agents actually smoother

OVHcloud Offices, Paris Tuesday, April 21

Ibrahim Guoual · MUG Paris

the problem

This is how most apps store chat history

messages = [] # The details are hidden. · can't query · can't debug

When your agent fails at 2am, what do you look at?

reality check

A conversation is not a list of strings

What you think it is

messages = [
  {"role": "user", "content": "what's the weather?"},
  {"role": "bot", "content": "it's 18°C in Paris"}
]

What it actually is

{
  role: "assistant",
  content: null,
  tool_calls: [{
    name: "get_weather",
    args: { city: "Paris" }
  }],
  usage: { tokens: 312 },
  ts: "2025-04-21T19:02Z"
}

Tool payloads differ per tool. MongoDB stores all shapes, no schema, no migration.

demo

Insert it. Then find the bug.

db.sessions.insertOne({
  session_id: "abc123",
  messages: [
    { role: "user", content: "Book a flight to Lyon" },
    { role: "assistant",
      tool_calls: [{ name: "search_flights",
                     args: { from:"CDG", to:"LYS" } }] },
    { role: "tool", name: "search_flights",
      result: { error: "API timeout", flights: null } }
  ]
})

// find every session where a tool failed
db.sessions.find({
  "messages.result.error": { $exists: true }
})

honest verdict

When MongoDB wins, and when it doesn't

Use MongoDB when

You need persistence across restarts
Multiple users and sessions
Agents that use tools
You want to debug failures
Analytics on agent behaviour

Redis is fine when

Last N messages only
Single session, no persistence
No need to query history
Raw speed is everything

"Storing history in memory is a choice.
Most people make it by accident."