personizeai/governed-memory

This repository contains the research materials, datasets, experiment protocols, and results, for a paper about governed memory, a system designed to give AI agents reliable, rule-following memory in business workflows. In the context of AI agents, memory refers to the ability to store and retrieve information across interactions, things like facts about a customer, decisions that were made, or context from previous conversations. The paper's argument is that this memory needs to be governed: it should follow business rules about what information can be shared, with whom, and under what conditions. For example, a sales agent should not be able to recall confidential competitor pricing when talking to a customer. The repository itself is not a runnable application. It is a collection of synthetic, entirely made up, no real people or companies, test documents and structured experiment data used to evaluate the performance of this kind of memory system. The synthetic documents include meeting transcripts, email threads, chat logs, account notes, and call notes. Each comes with a paired ground truth file specifying what facts should be extracted from it and what the correct answers to recall queries should be. The 15 experiments in the collection test different aspects: how accurately the system extracts facts from different content types, how well it deduplicates overlapping information, how reliably it enforces governance rules even when given adversarial inputs designed to bypass them, and how quickly it can retrieve relevant information. The materials are provided so that other researchers can reproduce the paper's results using the same API.