TypedMemory – long-term memory and reflection for AI agents (github.com)

TypedMemory, a new library for AI agents, introduces a long-term memory system designed to enhance agent behavior over time by facilitating the retention, recall, and reflection of information. Unlike traditional memory systems that simply store information, TypedMemory enables agents to surface contradictions and maintain an audit trail of changes, thus allowing for more reliable and reflective operations. This means agents can now keep track of evolving preferences, resolved goals, and historical context, addressing issues such as silent overwrites and forgotten decisions inherent in prior architectures. The technical architecture of TypedMemory is significant as it provides structured memory management, integrating a system of classification for different memory types, conflict policies, and an evolving framework through which information can be updated without losing the original data. This robustness is pivotal for applications in multi-tenant environments, legal and medical fields, and any setting where maintaining the integrity of information is crucial. Additionally, its ease of integration via Python and its lightweight dependencies position it as a compelling tool for AI/ML developers looking to create more intelligent and accountable agents that actively learn and adapt over time.