MemoryFlow: Real-Time, Implementation-Agnostic Telemetry for Measuring Dynamic Memory Quality in LLM Agents

MemoryFlow is a telemetry-only protocol specification for measuring dynamic memory quality in LLM-based agents.It defines deterministic event-stream semantics, audit-ready invariants, and fail-closed conformance profiles.It enables comparable, gaming-resistant metrics without prescribing any particular memory algorithm. This paper specifies a minimal event schema for memory operations (declare, write, replace, delete, read, use, verify, correct) and a deterministic verifier ordering for multi-collector observation. MemoryFlow treats observability as the primary boundary: only declared telemetry is measurable, and missing bindings or hidden rewrites downgrade or invalidate version-sensitive metrics under strict profiles. The specification introduces online-computable metrics derived from event streams with bounded verifier state, including churn (effective mutation rate), uptake (whether new writes are actually used), correction latency (time from failure to correction), verified/proven update fractions (VUF/PUF), and exposure measures for staleness, risk-weighted staleness, zombie reads/uses after deletion, and supersedence reads/uses after replacement. It also defines canonicalization and collision-robust version binding via content digests and update identifiers, enabling auditability and comparability across heterogeneous implementations. MemoryFlow is designed for practical integration with observability stacks, with informative guidance for mapping events into OpenTelemetry logs/spans and OpenInference tracing. The result is a protocol and metric family that separates “what is measurable from telemetry” from “how memory is implemented,” supporting evaluation, benchmarking, and telemetry contracts for real-world agent systems.