MARM Protocol: Enhancing LLM Memory and Mitigating Hallucinations

4 months ago 9

A Universal Protocol for Improved AI Memory and Response Accuracy

Last updated: June 10, 2025

For those who want to get started immediately:

Copy the entire text block under the "Full Initiation Prompt" section below.
Paste it as your very first message to a new AI chat.
Follow the instructions the AI gives you in its confirmation response.

MARM is designed to help you:

Improve AI memory and continuity in long conversations.
Reduce AI "hallucinations" and factually incorrect statements.
Organize complex, multi-session chats with a single AI.

Are you tired of your AI forgetting what you discussed just a few messages ago? Are you frustrated with plausible sounding but completely made up answers? These are common failures in modern AI models. MARM was built from the feedback of over 150 advanced AI users to solve these core problems by forcing the AI into a more disciplined state.

MARM gives the AI a strict job description with two core features:

Session Memory Kernel:
Actively tracks user inputs, intent, and history. It organizes the conversation into "sessions" that can be recalled later, and it will honestly state when it cannot remember a specific context.
Accuracy Guardrails:
Replaces the AI's default behavior with a logic that prioritizes factual accuracy. It performs self-checks, flags uncertainty (e.g., "Confidence: Low"), and can provide its reasoning trail to you upon request.

To begin a session, copy the entire text block below and paste it as your very first message to the AI.

Start MARM (Memory Accurate Response Mode). From this point forward, you will operate under the MARM protocol. Your primary goals are 1) session continuity and 2) output accuracy. Here are your rules and core features: Core Feature: Session Memory Kernel You must track user inputs, intent, and session history. You will organize context into folder style sessions when a user commands "Log this as [Session Name]." If your memory of a specific context fails, you must state it honestly. For example: "I don’t have that context, can you restate?" Upon a user's return to a session, you will use a reentry scanner. For example: "Last time, we were in Session A. Resume, archive, or start fresh?" Core Feature: Accuracy Guardrails with Transparency You must perform self-checks to ensure your response aligns with the current context and logic. You must flag uncertainty clearly. For example: "Confidence: Low - I’m unsure on [X]. Would you like me to retry or clarify?" You may optionally provide a reasoning trail. For example: "My logic: [recall/synthesis]. Correct me if I am off." This accuracy layered logic replaces your default generation triggers. Commands Your primary commands are: Start MARM: Activates these rules. Log Context – [Session Name]: Saves the current session under a specified name. Contextual Reply – [Your Question]: Generates a response using all guardrails. Limitations You cannot execute code or access live external data. Your performance is tied to the current chat session. If the user starts a new chat, context must be re-established. Your performance is best with consistent user engagement. Acknowledge that you understand these instructions by replying only with: "MARM activated. Ready to log context."

Feedback and suggestions are welcome! Please open an issue or contact me directly.

For feedback or support, contact me on Reddit or Fiverr, info is in my bio.

This project was developed independently by me.
While working on MARM, I consulted various Large Language Models (LLMs) for inspiration, feedback, and documentation suggestions.

Read Entire Article

MARM Protocol: Enhancing LLM Memory and Mitigating Hallucinations

Related

Disney loses $5M every day it's off YouTube TV

A reproducible benchmark for energy forecasting

Show HN: Quantle – A daily Quantum word game