EchoGuard: An Agentic Framework with Knowledge-Graph Memory for Detecting Manipulative Communication in Longitudinal Dialogue

EchoGuard is an agentic AI framework that utilizes a Knowledge Graph-based memory system to track longitudinal dialogue, detect psychologically grounded manipulative patterns, and guide users toward self-discovery through targeted Socratic prompts.

The Gist

Imagine you are walking through a dense, foggy forest. You know something feels "off" about the path you're on, but you can't quite see the traps. You feel confused, guilty, or anxious, but when you try to explain it to a friend, they say, "But they said they were just worried about you!"

This is what it feels like to be on the receiving end of manipulative communication (like gaslighting, guilt-tripping, or emotional blackmail). The problem isn't that you are crazy; it's that the "fog" (the manipulation) is designed to make you doubt your own senses.

The paper introduces EchoGuard, a new digital tool designed to act as a compass and a flashlight for people walking through this fog. Here is how it works, broken down into simple concepts:

1. The Problem: Why Current AI Fails

Think of current AI chatbots as having a very short-term memory. They can remember what you said in the last five minutes, but they forget what happened last week or last month.

Manipulation is like a slow leak in a boat. A single drop of water (one mean comment) isn't a problem. But over months, those drops add up, and the boat starts sinking. Because standard AI forgets the past, it can't see the "slow leak." It only sees the current drop of water and says, "That's not a big deal."

2. The Solution: EchoGuard's "Digital Scrapbook"

EchoGuard is different because it doesn't just chat; it builds a Knowledge Graph.

• The Analogy: Imagine a detective's corkboard.
  • The Old Way: The detective looks at one photo at a time.
  • EchoGuard's Way: The detective pins photos, connects them with red yarn, and draws circles around patterns.
  • How it works: When you log an interaction, EchoGuard doesn't just save the text. It breaks it down into "nodes" (dots) and "edges" (strings). It connects the Person to the Event, the Event to your Emotion, and the Emotion to a Manipulation Tactic.

Over time, this board fills up. Suddenly, the AI can see the red yarn connecting "Last Tuesday's guilt trip" to "Yesterday's apology" to "Today's self-doubt." It sees the pattern that you are too stressed to notice.

3. The Three-Step Dance (Log-Analyze-Reflect)

EchoGuard operates in a loop, like a helpful coach:

• Step 1: The Log (The Journal)
  Instead of just typing a rant, you answer a structured set of questions. "How did you feel?" "What exactly did they say?" "Did you feel confused?"

  • Why this helps: This forces your brain to slow down and label your feelings (a psychological trick called "affect labeling"). It turns a messy, scary emotion into clear data points for the corkboard.

• Step 2: The Analyze (The Detective)
  The AI looks at your "corkboard" and compares it to a library of known manipulation tactics (like a "Cheat Sheet" for bad behavior).

  • The Magic: It doesn't just look for bad words. It looks for context. It asks: "Hey, this person said 'I'm worried about you' three times this month, and every time, you felt small. Is this actually worry, or is it control?" It connects the dots that you missed.

• Step 3: The Reflect (The Socratic Coach)
  This is the most important part. EchoGuard never tells you what to do. It won't say, "Leave him!" or "He is an abuser."

  • The Analogy: Instead of driving the car for you, it sits in the passenger seat and asks, "Hey, notice how every time you say 'no', they get quiet and you feel guilty? What do you think is happening there?"
  • It uses Socratic questioning to help you realize the truth. It gives you the flashlight so you can see the trap, rather than just telling you where the trap is.

4. Why This Matters (Safety & Autonomy)

The creators are very careful. They know that if an AI tells a victim "Your partner is evil," the victim might get defensive or feel unsafe.

• The Guardrails: EchoGuard is designed to be a mirror, not a judge. It focuses on behaviors and patterns, not on labeling people as "bad."
• The Goal: The goal isn't to make the AI the hero; the goal is to help the human regain their own superpower: trust in their own intuition.

Summary

EchoGuard is like a smart, long-term memory assistant for your relationships.

• It remembers the history that your brain is too tired to track.
• It connects the dots to show you the "red flags" you missed.
• It gently asks questions to help you realize, "Oh, I'm not crazy. This is actually a pattern of manipulation."

It empowers you to see the fog clearing, so you can decide for yourself which path to take.

Technical Summary

Based on the paper provided, here is a detailed technical summary of EchoGuard, an agentic framework designed to detect manipulative communication in longitudinal dialogues.

1. Problem Statement

The paper addresses the critical challenge of manipulative communication (e.g., gaslighting, guilt-tripping, emotional coercion) in interpersonal relationships. Key issues identified include:

• The Awareness Gap: Individuals often experience emotional distress from manipulation but fail to consciously recognize the manipulative intent due to cognitive biases and the subtle, context-dependent nature of the tactics.
• Limitations of Existing AI: Current AI systems (sentiment analysis, toxicity classifiers) operate on single-utterance features and fail to capture relational dynamics or implicit harm. They lack the longitudinal memory required to detect patterns like intermittent reinforcement or escalating gaslighting over time.
• Technical Constraints: Relying solely on Large Language Models (LLMs) with expanding context windows is computationally inefficient, costly, and prone to "catastrophic forgetting" of earlier interaction details.
• Autonomy Concerns: Existing systems often aim for automatic classification rather than scaffolding user self-discovery, potentially removing user agency.

2. Methodology: The EchoGuard Framework

EchoGuard is a stateful agentic AI framework that bridges the gap between felt experience and conscious recognition using a Knowledge Graph (KG) as its core memory. It operates on a Log-Analyze-Reflect loop.

A. Core Architecture

The system utilizes a ReAct (Reasoning + Acting) agent coupled with a persistent Knowledge Graph ($G = (V, E)$) that serves as both episodic and semantic memory.

• Episodic Memory (User History): Stores nodes for User, Other_Person, Interaction_Event, Emotion, Tactic, and Phrase. Edges represent relationships like felt_emotion, used_tactic, and contains_phrase.
• Semantic Memory (Psychological Constructs): A pre-defined KG containing markers for six specific manipulation tactics (e.g., Gaslighting, Guilt Induction, Emotional Blackmail).

B. The Three-Phase Workflow

Phase 1: Structured Interaction Logging & Episodic KG Enrichment

• Input: Users log interactions via a structured questionnaire capturing communication content and emotional responses.
• Mechanism: The Structured Logger transforms subjective experiences into machine-readable nodes and edges.
  • It records Emotional States (using Plutchik's wheel) and Cognitive Assessments (e.g., self-doubt).
  • It creates a "computational signature" for the Awareness Gap by identifying events with high distress but low-confidence articulated causes.

Phase 2: Reasoning over the KG for Pattern Detection

• Mechanism: The Pattern Detection Engine performs a hybrid analysis:
  1. Graph-Based Reasoning: Executes multi-hop queries on the user's episodic KG to find subgraphs matching psychological markers (e.g., linking Interaction_Event $\to$ Emotion:Self-Doubt $\to$ Phrase:Reality_Denial).
  2. Semantic Analysis: Uses Sentence-BERT to match user phrases against expert-validated pattern banks.
• Output: Weighted confidence scores that validate the user's experience without issuing a definitive "verdict."

Phase 3: KG-Grounded Reflective Prompts

• Mechanism: The Prompt Generator uses an LLM to create interventions based on the retrieved subgraphs.
• Constraints: To prevent psychological reactance and preserve user agency, the system employs a three-layer constraint architecture:
  1. Template Constraints: Prohibits diagnostic labels (e.g., "abuser") and directive advice (e.g., "leave them").
  2. Grounding Strategy: The LLM is fed only the relevant, structured subgraph data (solving context window issues) rather than raw chat history.
  3. Adaptive Learning: The agent adjusts sensitivity based on user feedback to avoid rigid interpretations.
• Goal: Generate Socratic questions (e.g., "What assumptions about obligation did that phrase rely on?") to guide users toward self-discovery.

3. Key Contributions

1. Novel Agentic Framework: Introduces the Log-Analyze-Reflect loop, leveraging a Knowledge Graph as external episodic memory to model and detect "Awareness Gaps" in a computationally tractable way.
2. Hybrid Detection Method: Combines graph-based reasoning (for longitudinal pattern recognition) with semantic analysis (for linguistic nuance) to identify six categories of manipulative tactics.
3. Prototype Implementation: Demonstrates the feasibility of using KG-grounded, constrained LLM interventions for sensitive personal safety applications, prioritizing user autonomy over automated judgment.

4. Proposed Evaluation Strategy

The paper outlines a comprehensive evaluation plan (not yet completed in the text, but proposed):

• Experimental Design: A Multi-Arm Randomized Controlled Trial (RCT) comparing EchoGuard against:
  1. Structured Logging only.
  2. Psychoeducation Baseline.
  3. Control reflection task.
• Metrics: Primary outcome is Manipulation Recognition (identifying tactics and evidence). Secondary outcomes include metacognitive awareness and safety (anxiety levels).
• Baselines & Ablations: Comparisons against standard toxic language detectors, zero-shot LLM prompts, and agents using vector databases/flat logs to prove the superiority of KG-based reasoning.
• Failure Mode Analysis: Specifically testing for false positives (misidentifying assertive communication) and cultural context dependence.

5. Significance and Ethical Considerations

• Significance: EchoGuard represents a shift from content moderation (flagging toxic words) to cognitive scaffolding (helping users understand why they feel distressed). It addresses the "longitudinal memory" problem in agentic AI, offering a scalable solution for detecting subtle, evolving psychological abuse.
• Safety & Ethics:
  • Non-Diagnostic: The tool is explicitly educational, not diagnostic.
  • Agency Preservation: It avoids labeling people as "abusers" and refrains from giving directive advice, focusing instead on behavioral patterns.
  • Privacy: Designed as a first-person self-reporting tool to prevent misuse for surveillance; no personally identifiable information (PII) is stored.
  • Limitations: Acknowledges reliance on user-logged data (which may be incomplete) and the need for cross-cultural validation of the psychological constructs.

In summary, EchoGuard proposes a sophisticated architecture where Knowledge Graphs provide the necessary structural memory to detect complex, time-dependent manipulation patterns, while constrained LLMs act as empathetic guides to foster user self-awareness without compromising autonomy.