Mnemara – an LLM agent runtime where the role doc runs every turn (github.com)

Mnemara has launched as a new conversation runtime for Claude agents that allows for continuous enforcement of rules during every API call. Unlike traditional setups where system prompts are only applied at the start, Mnemara re-reads a role document (in Markdown format) at each turn, enabling the agent to self-monitor and prevent common pitfalls such as getting stuck on tasks, repetitive polling, drifting from user instructions, or instinctively changing conclusions based on user feedback. This is particularly significant as it enhances efficiency and reduces unnecessary costs associated with API usage. Developed on the Claude Agent SDK, Mnemara provides a straightforward setup process with powerful configurable options. Users can quickly initialize an agent, set rules, and run tasks with commands provided in its repository. The self-monitoring capabilities represented by the role document help refine the agent's behavior in real-time, ensuring it adheres to specified strategies for managing errors or failures. Mnemara is positioned to optimize how users interact with AI, encouraging responsible usage while maximizing productivity through customizable roles that are easily editable during sessions.