MemorySync
Getting Started

Respan Memory Observability

MemorySync memory + Respan (formerly Keywords AI) gateway observability: recall, LLM call, and persist as one fully traced, cost-tracked waterfall. Respan documents that the incumbent’s memory operations cannot be traced at the SDK level — MemorySync ships the first instrumentor that can.

The pattern: recall → gateway → persist

pip install opentelemetry-instrumentation-memorysync openai

Respan routes LLM traffic through one OpenAI-compatible gateway (https://api.respan.ai/api/ with your Respan key) and turns every call into traces, spend metrics, and evals. Respan’s tracing stack is OpenTelemetry underneath — the same pipeline this instrumentor emits into — so the memory spans and the gateway’s LLM spans land in one waterfall: you see the recall’s latency, the model call’s cost, and the persist’s idempotency signal side by side.

The gap this closes

Respan’s own Mem0 integration page states verbatim: *"Mem0 doesn’t expose its own SDK-level tracing instrumentor; route all calls through the Respan gateway to get full observability automatically."* Routing Mem0’s internal LLM calls through the gateway shows those LLM calls — but the memory operations themselves (the add, the search) never appear as spans. A slow search, a failed add, or a quota rejection is invisible.

Mem0 + RespanMemorySync + Respan
LLM calls traced✓ via gateway✓ via gateway
Memory operations as spans✗ invisible (no SDK instrumentor)✓ first-class memorysync.* spans
Recall latency visible✓ with server-side latency attribute
Failed persists visible✓ ERROR spans with HTTP status
Memory content in tracesn/anever, unless explicitly opted in

Supported versions

SurfaceRequiresVerified on
opentelemetry-instrumentation-memorysync 1.0.0Python 3.9+; any OpenAI-compatible SDK for the gateway side21 CI checks (shared with the AgentOps pairing): real-SDK spans, privacy default, pass-through fidelity, error spans, lifecycle, late-binding provider seam

Where to go next

Was this page helpful?