Wazari gives its AI agent a memory people can inspect, review and roll back

Wazari.ai says it built its Azure-based agent Ezra a memory system that people can inspect and change, with human review for proposed skills and checks for decisions whose assumptions have changed. It is a practical answer to a problem that arrives once an agent is expected to remember more than the last conversation.

Microsoft AI Watch analysis

What happened

Wazari stores each worker’s “brain”, including its memories, skills and reference material, in a GitHub repository. People can read the contents, compare changes and revert them. The company describes the approach in a Microsoft for Startups account of how it built Ezra.

A scheduled process called the dreamer reviews conversations and proposes memory changes. A separate process opens a pull request for each skill it seeds, leaving a person to decide whether it should be merged. Wazari also says it treats conversation transcripts as data, not instructions. Its open-source project kpopper records the premises behind conclusions and can flag a conclusion for review if one of those premises changes.

Why it matters

Persistent memory makes an agent more useful across sessions, but it also gives yesterday’s mistake a chance to become tomorrow’s confident answer. Wazari’s design makes the memory visible and reversible, and gives proposed skills a review gate rather than letting every learned procedure quietly become policy.

The record of assumptions tackles a subtler problem: a conclusion can become stale even when nobody edits the conclusion itself. In Wazari’s example, changing file retention from 30 days to seven means a decision to keep download links valid for 30 days needs another look.

Our read

This is a useful design pattern, not proof that every agent memory needs GitHub. The important choices are clearer than the storage brand: make saved knowledge inspectable, verify claimed actions somewhere other than the agent’s own account, and decide which changes need a human before they take effect. Memory without a way to correct it is just a very tidy route to repeating old mistakes.

What to watch

• Whether Wazari connects kpopper more closely with the memories used by its agents.
• How often human reviewers reject or revise proposed memories and skills.
• Whether the approach is practical for teams whose agent knowledge does not already live alongside code.

Discussion spark: Should persistent AI memory be managed like code, with visible history and human-reviewed changes, or does that add more process than most teams need?

not affiliated with or endorsed by Microsoft