Open Source
Tools that try to make AI accountable: from memory and verification to auditing model routing and protecting career decisions. All released under public licenses, with the code open to read.
Tools
Most of these tools began as a need: something missing on a research evening, written the next morning. I publish them as open source because the problems a psychologist meets while working with AI are not unique to me, and keeping a solution closed helps no one. Everything below can be read, used, and adapted to your own work.
Persistent memory
Most AI agents reset their memory every session. mneme closes that gap: a persistent memory system that stores context, recalls it across sessions, and shares it across multiple clients. I wrote it so a researcher's work does not start over with every session.
Explore the details onourimpram.github.io/mnemeVerification layer
AI agents write more and more code, yet it remains hard to know whether what they produce has actually been verified. mergen is a verification and governance layer that aims to make agent output accountable. It asks for the evidence behind a change, sets verification gates, and grounds trust in proof rather than claim.
Explore the details github.com/OnourImpram/mergenToolkit for social scientists
Social scientists now hold powerful AI tools, yet the path to using them within clinical and ethical bounds is rarely written down. I wrote this guide and toolkit for social scientists who want to work with AI while protecting data security and research integrity.
Explore the details github.com/OnourImpramZenodo DOIRouting audit
In an AI coding session it is rarely visible which model actually served each turn, and routing can drift from what you configured without anyone noticing. routeledger makes that drift visible by reading the session transcript itself. It is read-only and offline: nothing leaves your machine.
Explore the details onourimpram.github.io/routeledgerGitHub repositoryDecision safety
Career decisions are high-stakes, hard to reverse, and usually made alone. vocation-os was not written to hand those decisions to a model, but to keep them under human supervision. It gathers claims in a graph, puts gates in front of high-stakes steps, and writes every move to an append-only audit log.
Explore the details onourimpram.github.io/vocation-osGitHub repositoryThis list keeps growing. Some tools I am working on are still in private repositories, and they arrive here once they are ready. The most current view is always on my GitHub profile.
