Where Alyx works
What Alyx can do
Models and your own providers
Alyx runs on an Arize-managed model by default. To use your own, add the provider under Settings → Account Settings → Integrations and pick its model in Alyx’s model selector. Supported: OpenAI, Anthropic, Gemini, Azure OpenAI, Vertex AI, AWS Bedrock, NVIDIA NIM, and OpenAI-compatible custom endpoints. Only models with enough tool-use capability appear. See AI provider integrations.Settings
Open Settings from the menu at the top right of the Alyx chat:- Keyboard shortcut: any modifier plus a key, saved per browser.
- Auto Accept: skip confirmation for eval and task updates, annotations, datasets, prompt changes, and experiment runs. Labeling queues and custom views always ask.
- Profile: see add your own context.
- Documentation: allow the docs skill, which uses RunLLM. See third-party integrations.
- Tips: show or hide tips above the chat input.
Chat features
- Edit & resend a message to rewrite the conversation from that point. Alyx confirms first if the message created a dataset, evaluation, or experiment.
- Queue a message while Alyx is working. Send Now sends it right away.
- Progress shows a plan that updates as each task finishes.
- History on Alyx home reopens past conversations.
Long-term memory
Alyx remembers durable context between conversations, so you don’t re-explain the background: how you work, your goals and progress, what it learned about your projects and assets, and your team’s conventions and corrections. It keeps what stays useful, rather than a transcript or values it can look up again.Add your own context
- Tell it in chat. For example, “Remember that
support-regressionis our release-gating dataset.” What you state directly is stored as a high-confidence entry, and a correction replaces the stale one. Memory is written once the turn finishes, so it’s in place for your next message. - Fill in your profile under Settings → Profile: your name, your role, and instructions Alyx keeps in mind, then Save profile. Instructions are screened for safety when saved. A profile holds up to 10,000 characters, and roughly the first 8,000 go into each conversation, so lead with what matters most.
Long-term memory FAQ
Who can see what Alyx remembers about me?
Who can see what Alyx remembers about me?
Only you. Memory is stored per user, so a teammate in the same space builds their own memory and never reads yours. It is deleted along with your account or the space it belongs to.
Does memory follow me between spaces?
Does memory follow me between spaces?
No. Learned memory is scoped to the space where the conversation happened, so work in one space does not leak context into another. Your profile is the exception: it is tied to your user and applies everywhere you chat with Alyx.
Does Alyx keep a transcript of our conversations?
Does Alyx keep a transcript of our conversations?
No. Memory holds a short, curated set of notes rather than a log. An entry is kept only if it is still true next month, cannot be recovered with a fresh query, and shapes more than one request. Your conversation history is separate and available from Alyx home.
What does Alyx deliberately not remember?
What does Alyx deliberately not remember?
Anything it can look up again: evaluator and task configurations, IDs, URLs, filter strings, and live values such as today’s error count or current p99 latency. It does store results that mark progress toward a goal, like “prompt v4 reached 74% accuracy against a 90% target,” because later results get compared against them.
How do I remove or correct something Alyx remembers?
How do I remove or correct something Alyx remembers?
Tell Alyx in the chat. Say what is wrong or no longer relevant, such as “we are not tracking latency on that project anymore,” and it replaces or deletes the entry on the next update. There is no manual memory editor today, so corrections go through the conversation.
What if my profile and learned memory disagree?
What if my profile and learned memory disagree?
The profile wins, because it is the context you set yourself. If learned memory still carries something stale, say so in the chat as well so Alyx updates the entry instead of working around it.
Is my memory sent to the model provider?
Is my memory sent to the model provider?
Yes. Memory and your profile are included in the conversation like any other context, so they go to the model provider selected for that chat. See Data privacy for how Arize-managed and bring-your-own integrations handle that data.
Experience awareness
Alyx adapts its help to your familiarity with each part of Arize AX. Based on your experience level, Alyx can be an onboarding buddy that teaches as it works or an advanced, focused agent that concentrates on completing your tasks. The experience is catered to you and treated separately by product area, so Alyx can move quickly when you are working with tracing while providing more guidance when you try evaluations or dashboards for the first time.
Long-term memory and experience awareness work together: Alyx can remember your goals and preferences while adjusting how much guidance it provides in the part of AX you are using now.
Data privacy
Alyx sends requests to the model provider selected for your conversation. Arize-managed options include Azure OpenAI and Anthropic. With your own integration, that provider’s configuration and data-handling terms apply.- Azure OpenAI and Anthropic act as data processors for the prompts Alyx sends and the outputs it generates, depending on which Arize-hosted model is used, and help Arize meet industry-standard security and compliance measures. Processing and retention depend on the service and its configuration, and each provider describes its training and data-use commitments in its own documentation: see Azure OpenAI data, privacy, and security and Anthropic’s trust and privacy documentation.
- Azure OpenAI is fully controlled by Microsoft and hosted in Azure. It doesn’t interact with any other OpenAI-operated services, such as ChatGPT or the OpenAI API.
- Claude requests run on Anthropic infrastructure under Anthropic’s enterprise API security and data commitments, separate from unrelated consumer products.
