Heather Downing

Heather Downing is a developer advocate whose technical writing connects application development, identity security, and persistent context for AI agents. At the 2026 AI Engineer World’s Fair, she represented Yugabyte and demonstrated Meko, its agent-native persistence layer. Her examples address concrete questions: whether an application reliably records a decision, who can access it, and how a human correction should affect later work.

From hospitality management to developer advocacy

Downing entered software development after years in hospitality management. In 2011, she enrolled in a six-month Kansas City boot camp focused on ASP.NET application development. An internship at a mobile-app company gave her an early opportunity to build its service stack. She went on to develop enterprise services and proof-of-concept applications spanning APIs, mobile software, and voice interfaces.

She joined Okta’s developer-relations team in 2019, making developer education her full-time work. Her technical writing made identity integration concrete for C# developers through Alexa integrations, authenticated Blazor applications, and ASP.NET applications connected to hosted databases. In 2021, she was a principal developer advocate at Okta, a Microsoft MVP, and a .NET Foundation member. She also encouraged people with nontraditional backgrounds to enter software development—a path she had taken herself.

Authentication on a shared voice device

Downing’s work on voice interfaces examined the gap between connecting an account and verifying the person using it. Alexa account linking establishes an account connection, but does not establish who is speaking during a later interaction. On a shared household device, that distinction matters when a skill exposes financial information or permits purchases.

Her voice-only authentication example used .NET and Okta to add an interaction-level verification step without requiring the user to switch to a screen. It addressed both the security problem and the constraints of the interface: verification had to fit the way someone actually used a voice application.

Making agent memory dependable

In her Meko tutorials, Downing explains how persistent context becomes useful through reliable writes, selective sharing, and scoped corrections. The examples make those behaviors observable rather than leaving them to an instruction asking a model to remember.

• Deterministic context persistence: Required memory writes belong in application control flow. Giving a model a memory tool leaves it free to skip the call; an instruction to remember does not guarantee a record. Her persistence example wraps a Strands agent with direct calls to Meko, recording input before inference and results afterward. A separate audit process reads the stored trace back to check that the information survives beyond the original run.

• Governed agent handoffs: Saving information does not decide what another agent should receive. Her handoff example separates account-private memory from shared knowledge and uses a reviewable policy to promote selected decisions. A chef agent records menu choices, a kitchen agent develops purchasing notes, and a restaurant manager on another account receives only the promoted menu decisions. Unresolved questions and internal shopping details stay private, making both the transfer and its access limits visible.

• Human corrections that carry forward: Her Jev code-review example addresses repeatedly dismissing the same automated finding. It stores approved reviewer rulings outside the model and retrieves them when relevant. The check first evaluates a change against the guidelines, then separately determines whether a previous ruling exempts it. A fixture-specific exception includes its scope and reason, helping prevent that exception from spreading to production code.

Her World’s Fair Meko demonstration brought persistence and sharing together: resuming context in a fresh session, passing it between agents, and promoting private memories into shared data packs. She presented alongside Yugabyte co-founder Karthik Ranganathan, who developed the talk’s argument about the difference between individual agent memory and collective learning.