---
title: "Cadence Cue"
description: "A local memory system that gives your coding agent a starting guide - relevant files, tests, and gotchas - before it starts a task."
url: "https://teamcadence.ai/docs/cue/overview/"
---
## What Cue does

Cadence Cue turns your own session history into a resource your coding agent can consult *before* starting a new task, rather than starting cold every time. Ask it about a task, and it returns a short, evidence-backed starting guide: relevant files, relevant tests, and historical cautions - things that have bitten this codebase before.

```bash
cadence cue "add rate limiting to the webhook endpoint"
```

If Cue doesn't have reliable evidence for a task, it abstains rather than guessing. It's meant to be trustworthy evidence, not a general-purpose assistant filling gaps with plausible-sounding text.

## Building Cue's memory

Cue's memory is a local artifact - embeddings plus a graph - built from session history for a specific repo. Build or refresh it with:

```bash
cadence cue build --repo-path .
```

The first run downloads local embedding models (BGE/Jina-family) to your machine. After that, `cue build` reruns incrementally; `--force` rebuilds from scratch. `--max-sessions` and `--dense-cap` bound how much history goes into the artifact, and `--output-dir` controls where it's written.

## Exploring what Cue knows

`cadence cue browse` generates an interactive local HTML graph of your project's history, so you can see what evidence Cue is actually drawing on rather than treating it as a black box:

```bash
cadence cue browse --open
```

## Why it's local and opt-in

Cue is deliberately separate from the automatic session-upload path described in [Installation & onboarding](/docs/installation/). Nothing about Cue requires the desktop daemon to be running, and its memory artifact stays on your machine after the initial build - it isn't a new upload mechanism. Where Cue needs source material beyond your own local history (for example, org-wide evidence), it uses the same author-free, server-redacted export as `cadence session export` - see [Privacy & data](/docs/privacy/).

## How this differs from the CLI's other commands

Most `cadence` CLI commands (`session`, `review`, `blame`, `insights`) are read-only lookups against data Cadence already has server-side. Cue is the exception: it builds and queries a standalone local artifact. See the [command reference](/docs/cli/commands/) for the full list, and the [CLI overview](/docs/cli/overview/) for how agents are expected to use the CLI generally.