Experimental Evaluation

Dynamic STM and LTM in MineAI

Design, testing, and evaluation of MineAI’s Short-Term Memory (STM) and Long-Term Memory (LTM) systems.

Experimental Evaluation of Dynamic STM and LTM

Authors: MineAI Team

Publish Date: 21 December 2025

Abstract

This document presents the design, testing, and evaluation of MineAI’s Short-Term Memory (STM) and Long-Term Memory (LTM) systems. It describes how user messages are captured, stored, and retrieved dynamically, and explains memory behavior using a Redux-inspired pattern, highlighting state management, actions, and selectors. The LTM is currently in Experimental Phase, with plans to refine and launch as Beta.

1. Introduction

STM (Short-Term Memory)

Captures temporary, session-specific information. Resets after the session ends.

LTM (Long-Term Memory)

Stores persistent, cross-session user information and evolves dynamically based on frequency and importance.

The system follows a state-management approach similar to Redux, where actions update memory state, and selectors retrieve relevant information for personalized AI responses.

2. Memory Architecture

2.1 Short-Term Memory (STM)

StateStores all session-specific messages and context.
ActionsAdd new STM entry; Clear STM on session end.
SelectorsRetrieve session-specific info for AI responses.

Behavior: STM captures conversational context and ephemeral instructions, such as “respond using only emojis” or session-specific project mentions.

2.2 Long-Term Memory (LTM)

StateStores persistent user information across sessions.
ActionsExtract, Update, Decay, and Delete memories.
SelectorsRetrieve relevant LTM entries based on context.
DynamicsReinforcement and decay based on frequency.

Categories of LTM Memories:

PreferencesProjectsGoalsConstraintsFacts

3. Testing Methodology

3.1 STM Testing

Focus: Verify session-limited memory capture and retrieval.

Session-specific project name retrieval
Style-specific response enforcement
Automatic reset after session termination

3.2 LTM Testing

Category	Session 1 Input	Session 2 Retrieval
Preference	User prefers short and direct answers.	How should you reply? → Short and direct
Project	User is working on a hosting platform called SkyHost.	What project am I working on? → SkyHost
Goal	User’s goal is to become a leading AI hosting provider.	What is my long-term goal? → Become leading AI hosting provider
Constraint	User needs all responses to be concise.	How should you respond? → Enforce concise responses
Fact	User lives in Pakistan and is named Ali.	What’s my name and where do I live? → Ali, Pakistan

4. Memory Dynamics

Frequency-based reinforcement: Repeated mentions increase memory confidence.
Confidence decay: Unused or low-priority memories gradually lose strength.
STM vs LTM separation: Clear boundaries between ephemeral and persistent data.
Dynamic UI: Manual control via Settings → History Management.

5. Experimental Observations

SkyHost project persisted due to multiple mentions.
Single-mention facts disappeared if sessions were removed.
Cross-session retrieval works best for high-confidence topics.
System reflects human-like memory behavior.

6. Future Work

Fine-tune confidence decay and reinforcement parameters.

Ensure single-mention important facts persist (goals, constraints).

Expand memory extraction patterns for broader coverage.

Improve cross-session retrieval accuracy and History UI.