Respan + MemorySync

Overview

Respan routes LLM traffic through one OpenAI-compatible gateway (api.respan.ai) and turns every call into traces, spend metrics, and evals. Pair it with MemorySync: recall context before the gateway call, persist the turn after it, and instrument the memory side with opentelemetry-instrumentation-memorysync — Respan’s own tracing SDK is OpenTelemetry underneath, so memory operations appear as first-class spans next to the gateway’s LLM spans. That closes the gap Respan documents for the incumbent: their Mem0 page routes Mem0’s internal LLM calls through the gateway, but the memory operations themselves — the add, the search — are invisible as spans, and their docs state verbatim that "Mem0 doesn’t expose its own SDK-level tracing instrumentor." MemorySync does — it is the first memory vendor that ships one. The result: cost-optimized, fully tracked applications where a slow recall, a failed persist, or a memory quota event is visible in the same waterfall as the model call it affected.

Setup status and requirements

Supported version

opentelemetry-instrumentation-memorysync 1.0.0 (PyPI)

Last setup review

2026-08-25

Permissions

A MemorySync API key; a Respan API key for the gateway and dashboard.

Limits

Python 3.9+. The gateway pattern works with any OpenAI-compatible SDK; Respan was formerly Keywords AI and old keywordsai.co URLs redirect.

Capabilities

• Memory operations as first-class spans next to Respan gateway LLM spans
• Closes the incumbent gap: Respan documents Mem0 memory ops as un-traceable
• One gateway for routing + spend tracking; one line for memory tracing
• OpenTelemetry end to end — Respan’s tracing SDK and ours share the pipeline
• Privacy-first: memory content never recorded unless opted in, then truncated
• Recall-before / persist-after pattern documented with runnable code
• Failure visibility: error spans with HTTP status and error type
• Works identically if you later swap gateways — nothing is Respan-locked

Quick Start

from openai import OpenAI
from memorysync import MemorySyncClient
from opentelemetry.instrumentation.memorysync import instrument_memorysync

instrument_memorysync()
ms = MemorySyncClient(api_key="ms_...", base_url="https://api.memorysync.io")
llm = OpenAI(api_key=RESPAN_API_KEY, base_url="https://api.respan.ai/api/")

ctx = ms.recall(tenant_id=t, user_id=u, prompt=question, k=6)   # traced
reply = llm.chat.completions.create(model="gpt-4.1-nano", messages=[
    {"role": "system", "content": f"Memory:\n{ctx.get('context', '')}"},
    {"role": "user", "content": question},
])                                                              # gateway-traced
ms.add_turn(tenant_id=t, user_id=u, text=f"user: {question}")   # traced

Use Cases

Explore More