Summary Wazari.ai gave Ezra, its Azure expert agent, a memory a person can read and a review gate for proposed skills. It also built a separate record to flag conclusions when recorded premises change.
When people and agents work across many sessions, later conversations can lose sight of earlier decisions. Wazari.ai ran into that problem while building Ezra, a Microsoft Azure expert agent on its m8t platform. The team chose to keep the agent’s memory in a GitHub repository, maintain it through scheduled processes, and record the reasoning behind important conclusions.
Wazari.ai’s approach raises three practical questions for any team building agents: Can a person verify what the agent says it saved? Who reviews what it learns? And how will the team know when a conclusion needs another look?
Building an AI agent of your own? Microsoft for Startups helps founders build fast, scale smart, and sell more with Startup credits, AI tools, technical guidance, and go-to-market support.
Ezra’s architecture: Microsoft Foundry and m8t
Ezra runs as a prompt agent on Microsoft Foundry. Wazari.ai’s m8t platform provides its memory, tools, and execution layer; the execution and coding agents are containerized agents built on Microsoft Agent Framework.
When Ezra is installed in an organization’s Azure subscription, its “brain”—a repository of persona, memories, skills, and reference material—is forked into that organization’s GitHub organization.
Keep memory where a person can inspect it
Each worker’s memory, which Wazari.ai calls a brain, lives in a GitHub repository. Its owner can read, compare, and revert changes.
That history proved useful when a worker reported that it had saved a file, including a commit message and path, but the repository contained no corresponding write. Wazari.ai responded by treating the agent’s report as a claim to verify rather than proof that the action occurred. A save needs confirmation from something other than the agent; an unconfirmed claim of a write triggers a warning in the reply.
Review also surfaced an older Microsoft Foundry model and price saved as memory. A reviewer superseded the entry, leaving its history available. When two people supplied conflicting configuration details in different sessions, the team retracted one of the resulting memories.
Builder takeaway: Give people a way to verify what an agent actually wrote, inspect changes to persistent knowledge, and reverse those changes when necessary.
We chose GitHub because it already held everything our customers trust with their code, and a worker’s memory deserves the same treatment as code: a history, a diff, a reviewer, and a revert. A database would have been faster to query and impossible for a user to read without a tool.
Ilan Bar-Magen, Co-founder and CEO
Wazari.ai has also made Ezra’s brain template and public brain available as open source. Corrections to the public Ezra brain can be proposed through pull requests; an installed brain lives in the customer’s GitHub organization.
Separate learning from promotion
Wazari.ai uses two scheduled processes to maintain a worker’s brain. The first process, which Wazari.ai calls the dreamer, reviews conversations and proposes memory changes. Those proposals can include adding a fact, superseding or retracting an existing one, seeding a skill, or quarantining material that looks like an instruction. Platform code applies its consolidation changes and marks their origin, so a person can distinguish them from other changes. Conversation transcripts are treated as data, not instructions.
A second process, the librarian, runs as a GitHub Action. One job maintains indexes and flags items on a do-not-capture list. Another takes skills seeded by the dreamer and opens a pull request for each one. A person decides what merges.
Consider a user describing their team’s GPU-quota request process. The dreamer could propose the region and ticket-queue details as memory changes. Turning the procedure into a reusable skill would require a pull request and human review.
Builder takeaway: Define which agent-generated changes can be applied by a process, which are proposals, and which require a person’s review.
Want to explore more GitHub developer content? GitHub Universe 2026 takes place October 28–29, with free virtual attendance.
Record what a conclusion depends on
Memory can outlast the reasons behind a decision. To help revisit decisions, Wazari.ai built kpopper, an open-source knowledge record that can be used independently of m8t. A kpopper project record connects sourced facts and rules to the conclusions that depend on them. It also records what would invalidate a conclusion. When a premise changes, the system can identify dependent conclusions for review.
In one of the team’s projects, a decision allowed download links to remain valid for 30 days, with both file retention and link validity values being set to 30 days. kpopper captured those assumptions alongside the decision. When file retention later dropped to seven days, the earlier conclusion was flagged for review because a file could now disappear before its link expired.
Builder takeaway: Don’t just record a decision. Record the assumptions behind it and the conditions that should trigger another review. Wazari.ai plans to connect this record more closely with promoted memories, so those memories carry both what a worker knows and what it rests on.
Want to explore the approach yourself? Try Ezra with an Azure question, or explore kpopper on GitHub to see how Wazari.ai records facts, assumptions, and conclusions.
Building agents with Microsoft for Startups
If you’re designing an agent of your own, start with the three review points: where its knowledge lives, how writes are confirmed, and when a learned skill or conclusion needs another look. For a broader look at agent context and orchestration choices, see Building your Agent on Azure.
Microsoft for Startups helps founders build fast, scale smart, and sell more with Startup credits, AI tools, technical guidance, and go-to-market support. If you’re building an agent or AI application on Azure, apply to get started with Startup credits and unlock more as you build and scale.
Access your startups benefits today
Microsoft for Startups helps founders build fast, scale smart, and sell more. Apply today to unlock up to $150,000 in Startup credits to start building immediately.
