This describes a production system that pipes Langfuse observability traces into Memanto agent memory, with the core insight that naive integration (firehose all traces) destroys recall by drowning the agent in near-duplicates. Instead, the system groups observations by error signature—normalizing volatile details like IDs and file paths—so 812 failing traces collapse into perhaps two distinct lessons, each with confidence scores that rise logarithmically with occurrence count. You get two ingestion paths: a live span processor (two lines of Python, zero API overhead) that catches errors and latency anomalies in real time, and a CLI sync that reads Langfuse server-side to capture scores and costs that spans can't see. The ledger deduplicates across both paths, the buffer is bounded, failures don't crash your app, and everything traces back to a representative Langfuse span. Result: before your agent plans its next move on, say, `generate-response`, it can ask what actually broke on that operation in production and avoid the approach that failed 812 times. The source is concrete about what each capture mode can and cannot observe live (scores require the sync, percentiles require a population), and explicit about failure modes—missing thresholds log warnings, not silent zeros.