The VENETA memory layer, open source.

For anyone running a local model. Every lesson the model learns is proposed, measured, approved by a person and written to a ledger. Apache-2.0. v0.1 on 2026-11-24.

Follow releases Changelog veneta Enterprise

What the project publishes

Quickstart

Run it beside a local model.

Integrations

The agent frameworks it sits under: Hermes Agent, OpenClaw, Paperclip, and any OpenAI-compatible endpoint.

Concepts

Rules, episodes, facts, outcomes, preferences, and the approval loop.

Policy

What is proposed, what is blocked, what is ledgered.

Benchmarks

The memory benchmark, veneta-bench, and how to rerun it.

Changelog

Every release.

Built for one operator

A person is the last step

A lesson the loop distils waits until you approve it; a rejected lesson is never proposed again.

Everything on a ledger

Every proposal, approval, rejection, block and undo is one line in an append-only ledger.

Telemetry off by default

Zero content bytes. Five tiers you choose, and you can show, switch off or purge it at any time.

Apache-2.0

The code is open; the veneta name and mascot follow a trademark policy.

Where it sits

Between your agent and your model. Point the agent at the veneta proxy instead of the model; veneta recalls memory into every turn and distils lessons after it. It is not an agent and not a model.

The same core runs inside VENETA Enterprise and the VENETA WorldModel products. Nothing that ships here ever moves to paid.