Skip to content

Observability (Logfire)

memFrame can send the entire agent pipeline to Logfire — every agent run, LLM call, and tool call, plus a top-level trace per request and spans around the guardrail, planner, and sub-query execution. This is opt-in and fail-open: if the logfire extra isn't installed, the token is missing, or configuration errors, logging just continues normally and your chat is never blocked.

Install the extra (it also includes the [ai] agent runtime, so this single command is enough to trace the agents):

pip install "memframe[logfire]"
# or
uv add "memframe[logfire]"

memFrame uses a Bring-Your-Own-Key (BYOK) model: you supply your own Logfire token and telemetry goes to your project — nothing is hardcoded or committed. Install the extra, then enable it when you turn the agent on (pass a token, or set the LOGFIRE_TOKEN environment variable):

pip install "memframe[logfire]"
# or: uv add "memframe[logfire]"   (includes the [ai] agent runtime and logfire[system-metrics])

await mf.aenable_agent(
    provider="openai",
    model="gpt-5.5",
    api_key="sk-...",
    logfire_enabled=True,
    logfire_token="<your-logfire-token>",   # or rely on LOGFIRE_TOKEN env
    logfire_environment="dev",
)

Local dev without a token (optional, not committed): after uv run logfire auth (browser OAuth) you can bind this repo to a project with uv run logfire projects use --org <your-org> <your-project>; then logfire.configure() needs no token. Do not commit the resulting [tool.logfire] binding — it pins the repo to a personal project. For CI / headless runs, set LOGFIRE_TOKEN instead.

What gets traced automatically:

  • instrument_pydantic_ai() traces every Pydantic-AI agent — the guardrail, the planner, each specialist, and the dashboard designer — with child spans for each LLM request/response and tool call.
  • instrument_logging() (newer Logfire) forwards the standard memframe.ai log lines (model requests, tool calls, guardrail verdicts) into Logfire as logs.
  • instrument_system_metrics() (via the system-metrics extra) populates the Logfire Hosts view (CPU / memory / disk) for the process.
  • An adashboard / achat span wraps each user request, with nested guardrail.verify, planner.plan, and execute_subqueries spans — so the whole flow shows as one coherent trace in the Logfire UI.

Open https://logfire.pydantic.dev/ to explore the traces.