A little context.
A useful head start.

01 / THE IDEA

Give the next task
a better starting point.

Lavett aims to reduce the context gathering and repeated discovery that consume AI usage. It stores eligible user prompts and selected files locally, then supplies relevant earlier text to a supported coding client.

The client can inspect an exact-source receipt. You can check the original words, see what was recalled, and decide whether it is relevant. The stored memory is evidence, not a new authority over your instructions.

Memory is the mechanism; the goal is more useful work for the same budget. Read the cost, context, tool-call and timing evidence.

02 / YOUR WORKFLOW

A memory outside
a single conversation.

Claude Code and Codex are the main onboarding path. Supported local hooks can capture eligible prompts and provide cited context in later sessions of the same project. Your chosen model stays in your chosen client.

Other integrations have different limits. Gemini, Cursor and VS Code have prompt-hook adapters; Claude Desktop and Windsurf use model-directed MCP tools. An adapter’s presence does not mean universal support or guaranteed activation. Native trust and the client’s own permissions still apply.

03 / GETTING STARTED

Currently in pre-release.

Lavett’s public download is not available yet. Local engineering and package checks do not clear its publication gates. If you already have an authorized source checkout, the integration path begins here.

Requires CPython 3.11.4 or newer in supported Python 3.11–3.14 releases, and uv.

1. Run from the source checkout. Review the proposed configuration changes and answer the installer’s confirmation.
2. Restart your client and complete its native trust steps. In Codex, hooks also need trust in /hooks.
3. Run uv run --extra agents lavett integrate doctor. Installation alone does not prove that the running client activated the hooks.

New to the project? Contact us about access and readiness.

04 / YOUR DATA

Local processing.
Explicit boundaries.

The default memory path runs locally and needs no model API key. Stores are scoped to workspaces and are not encrypted. Default capture records eligible user prompts; model answers are not treated as user evidence.

Recalled text enters your chosen AI client and may reach its model provider under that client’s policy. Installing packages can use the network. Optional remote providers are separate, explicit features with their own controls.

This website itself has no analytics, remote fonts, account system or data collection form. Its demonstrations run in your browser and do not capture your text.

05 / YOUR CONTROLS

Inspect. Pause.
Pick it up later.

Retraction, erasure, capture exclusions and store selection have separate rules. Consult the source checkout’s memory and retraction documentation before deleting data.

06 / KNOW THE LIMITS

Evidence you can inspect.
Judgment you keep.

Recall is lexical. It can miss paraphrases, omit relevant evidence or surface old assumptions. A source receipt establishes occurrence, not truth or correct use by the model. The broader engine also includes Current and adviser-review evidence packets; their individual gates and limits remain in force.

Reduced context is not automatically reduced cost. Actual bill or allowance reductions and a general hallucination-reduction rate remain unestablished. The results page includes the promising measurements, uncertainty and losses.