> ## Documentation Index
> Fetch the complete documentation index at: https://docs.negentro.tech/llms.txt
> Use this file to discover all available pages before exploring further.

# Understand how your memory system performs

> Measure the quality, speed, and cost of memory and retrieval workflows so you can improve the experience of your AI application over time.

## What to measure

> **Retrieval quality**
>
> Understand whether the system returns relevant and useful information for representative queries. Use labeled examples and task-specific evaluations to assess result quality.

> **Latency**
>
> Track end-to-end response time and, where instrumentation permits, separate the time spent in retrieval, context assembly, storage operations, and downstream model calls.

> **Usage**
>
> Measure request volume, records stored, retrieval frequency, context size, and other usage dimensions exposed by your deployment.

> **Reliability**
>
> Monitor errors, timeouts, retries, empty results, and dependency failures. Segment failures by operation and environment to identify recurring issues.

> **Cost and efficiency**
>
> Where cost data is available, evaluate the relationship between retrieval quality, token usage, infrastructure consumption, and overall application cost.

## Suggested dashboard layout

### Overview metrics

## Evaluate quality, not just activity

A high number of retrievals does not necessarily mean retrieval is working well. Build a representative evaluation set from your real use cases and measure whether the system returns the information needed to complete each task.

Review false positives, missed information, stale records, and conflicting results. Use the findings to improve data quality, metadata, retrieval settings, and context selection.

## Monitor latency end to end

Measure the full request lifecycle as well as individual stages. A slow response may originate in retrieval, storage, a model provider, or another dependency.

Track median and tail latency rather than relying only on averages. Set alerts based on the needs of your application and establish a clear process for investigating regressions.

## Observability and privacy

Logs and analytics should help diagnose system behavior without unnecessarily exposing user content. Prefer structured events, operation identifiers, latency measurements, and carefully controlled diagnostic samples.

Apply access controls and retention policies to telemetry. Redact sensitive values where appropriate and document which metrics are collected, where they are stored, and who can access them.

## From measurement to improvement


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.