AI engineering rarely happens in one conversation.
You might investigate a production issue today, return later to build an eval, compare an experiment, or refine a prompt. Along the way, you establish context that matters the next time you work.
Alyx, the AI engineering agent in Arize AX, can now retain that context across sessions with long-term memory. It can remember durable information such as project goals, conventions, preferences, and previous decisions, so you spend less time re-explaining your environment and more time continuing the work.
TL;DR: Alyx now includes’s long-term memory that keeps useful context across sessions and is enabled by default for all users of Arize AX. Memories aren’t shared with teammates or across spaces, and you can ask Alyx to show, correct, or delete them in chat.
Build better agents with Arize
Trace, evaluate, and learn. Build agents that work with Arize AX and start tracing your runs today.
Prefer open source? Try Arize Phoenix for self-hosted, open source agent observability.
What Alyx remembers across sessions #
Alyx can remember things like:
- how you define a successful response for your agent
- where ground truth lives
- a production target you are working toward
- dataset or workflow conventions
- approaches you rejected and why
The focus is on context that is useful across sessions and difficult to reconstruct from the platform alone.
Alyx can look up current configurations, IDs, and live metrics when it needs them, so storing those details as memories could mean relying on information that’s already out of date. That leaves memory for the context you shouldn’t have to explain again, such as why you chose a particular approach or which tradeoffs matter and why.
Find out how we built long-term memory in Alyx
Explore our guide for a full breakdown of the engineering behind long-term memory in Alyx (plus the considerations and choices we made).
Use remembered context across traces, evals, and experiments in Arize AX #
Alyx is available across traces, eval workflows, experiments, and other AI engineering surfaces in Arize AX.
For example, while investigating a support agent’s traces, you might explain that refund requests require a policy check before the agent recommends an action. When you return in a new session to build an evaluator, Alyx can use that remembered requirement to help define what the evaluator should check. Memory is scoped by user and Arize space, so context stays relevant to the work you are doing.
How we built long-term memory for Alyx #
Building useful memory means deciding what Alyx should remember, what it can retrieve from AX, and how that context should carry across sessions without getting in the way.
For a deeper look at the architecture, tradeoffs, and evaluation behind the system, read Priyan Jindal’s How we built long-term memory for Alyx.
Keep working without starting over #
Long-term memory gives Alyx more continuity across sessions and workflows, so you can return to a project with more of the important context intact.