aema

Adaptive AI Agent with Persistent Experience Memory

GenAI OJT Capstone Project — Group G146

---

Project Overview

Adaptive AI Agent with Persistent Experience Memory is an Agentic AI project exploring how an AI agent can retain, evaluate, and reuse operational problem-solving experience across repeated task executions.

While foundational Large Language Models (LLMs) solve individual tasks with high proficiency, multi-step agentic workflows reveal two critical vulnerabilities:

1. Experience Amnesia: A frozen model starts completely stateless in every new session, unable to recall prior problem-solving discoveries or avoid repeating past mistakes.
2. Negative Transfer: Naive retrieval of past conversational history often retrieves irrelevant context, flawed logic, or anti-patterns that degrade downstream task performance.

Our architecture solves this externally without fine-tuning or modifying base LLM weights. We introduce an external experience-memory intelligence layer that distills generalizable lessons ({Trigger, Strategy, Pitfall}), verifies empirical reliability using an evidence-based Trust Score ($S \in [0.0, 1.0]$), and dynamically augments prompts with verified strategies and negative constraints.

---

Research Question

Under what conditions does persistent experience memory improve future task performance, and can reliability-aware memory selection reduce the harm caused by misleading or outdated experiences?

---

Core Idea

The agent operates across a closed-loop cognitive lifecycle:

Task Input
    ↓
Retrieve Relevant Experiences (Composite Ranking: 0.70*Similarity + 0.30*Trust)
    ↓
Execute Task (Prompt augmented with positive Strategy + negative Pitfall)
    ↓
Evaluate Outcome (Deterministic scalar reward R ∈ [0.0, 1.0])
    ↓
Reflect & Extract Lesson (Structured 3-tuple: {Trigger, Strategy, Pitfall})
    ↓
Update Trust Score (Exponential Moving Average & Quarantine if S < 0.35)
    ↓
Persist to Vector Store (Supabase PostgreSQL + pgvector HNSW index)
    ↓
Reuse in Future Tasks

The system learns continuously from task execution without altering the underlying LLM. In this cognitive loop, experiences are not only retrieved based on semantic relevance; their empirical reliability is continuously scored and updated based on observed usefulness over time.

---

Why This Project

Standard RAG architectures retrieve static document passages for factual question answering. However, autonomous agents require Experience-RAG: remembering how to solve problems, which tools to invoke, and what mistakes to avoid.

Key design principles:

• Preserving Discoveries: Useful problem-solving strategies must persist beyond an individual session boundary.
• Relevance ≠ Reliability: Just because a memory is semantically similar does not mean it was historically successful.
• Active Quarantine: Unreliable or outdated memories must decay and be isolated ($S < 0.35$) before they trigger negative transfer.

---

Existing Work We Studied

We analyzed prior agent memory architectures to ground our research methodology:

• Reflexion (Shinn et al., 2023): Linguistic self-reflection within a single task episode, but lacked cross-session vector persistence and reliability decay.
• Voyager (Wang et al., 2023): Open-ended skill libraries in Minecraft, but assumed non-decaying deterministic code skills without negative constraint learning.
• Generative Agents (Park et al., 2023) & MemoryBank: Memory importance and recency decay for conversational personas, but did not focus on empirical task success reliability.
• Mem0 / Zep / Letta: Commercial memory stores focused on entity extraction and conversational key-value facts rather than structured strategy-pitfall pairs.
• ExpeL (Zhao et al., 2024): Highlighted the critical risk of negative transfer when agents blindly retrieve unverified past experiences.

Our architecture directly addresses the limitation identified in ExpeL by establishing an evidence-grounded Trust Score ($S$) and bilateral prompt augmentation (positive strategies + negative pitfalls).

---

Key Implemented Features

1. Interactive 3D Neural Memory Core

• Three.js Spatial Visualizer: An interactive 3D particle graph (ThreeMemoryCore.tsx) rendering memory nodes clustered by domain and color-coded by trust score ($S \ge 0.70$ emerald, $0.35 \le S < 0.70$ amber, $S < 0.35$ crimson).
• Domain Filtering: Real-time exploration across research, coding, reasoning, and general domains.

2. Live Supabase Authentication & Session Management

• GoTrue Engine: Secure user registration, password hashing (bcrypt), and JWT session persistence.
• Row-Level Security (RLS): Database-level multi-tenant isolation ensuring users only access their own task runs and memories.

3. Central Agent Dashboard (/dashboard)

• Telemetry Overview: Real-time telemetry monitoring active memories, average trust scores, success rates, and execution latencies.
• Task Runner: Multi-line prompt input with domain selection and memory retrieval inspection.

4. FastAPI Backend Gateway & Deterministic Health Check

• High-performance asynchronous API gateway running on Uvicorn.
• Native /api/v1/health and /health endpoints with automated pytest verification.

---

System Architecture

React 18 Dashboard ──(HTTP / Bearer JWT)──► FastAPI Gateway (/api/v1)
                                                    │
                                                    ▼
                                           LangGraph StateGraph
                                       ┌────────────┴────────────┐
                                       ▼                         ▼
                            LiteLLM Gateway              Experience Engine
                     (Groq / Gemini / Ollama)       (PostgreSQL 15 + pgvector)
                                       │                         │
                                       ▼                         ▼
                                 External Tools            HNSW Vector Index
                             (DuckDuckGo / Reader)      (Composite Retrieval RPC)

Evaluation and observability operate across the execution pipeline via LangSmith and immutable PostgreSQL audit ledgers (trust_history).

---

Technology Stack

---

Quick Start & Local Setup

Prerequisites

• Node.js 18.0 or later
• Python 3.9 or later
• npm 9.0 or later
• Git

---

1. Clone the Repository

git clone https://github.com/Sumit-ai-dev/Adaptive-Agent-Memory-OJT-G146.git
cd Adaptive-Agent-Memory-OJT-G146

---

2. Frontend Setup & Execution

1. Navigate to the frontend directory and install dependencies:

   cd frontend
   npm install
   

2. Configure environment variables:

   cp .env.example .env
   

   Edit .env with your Supabase credentials (or use the built-in mock fallback for instant offline preview):

   VITE_SUPABASE_URL=https://your-project-ref.supabase.co
   VITE_SUPABASE_ANON_KEY=your-supabase-publishable-key
   VITE_API_URL=http://localhost:8000
   VITE_ENABLE_MOCK_FALLBACK=true
   

3. Launch the development server:

   npm run dev
   

   Open http://localhost:5173 in your browser.

4. Verify type-checking and production build:

   npx tsc --noEmit
   npm run build
   

---

3. Backend Setup & Execution

1. In a separate terminal, navigate to the backend directory:

   cd backend
   

2. (Recommended) Create and activate a Python virtual environment:

   python3 -m venv venv
   source venv/bin/activate
   

3. Install backend dependencies:

   pip install -r requirements.txt
   

4. Start the FastAPI development server:

   uvicorn app.main:app --reload --port 8000
   

   Interactive API documentation (Swagger UI) is available at http://localhost:8000/docs.

---

4. Running Automated Tests

From the repository root (return to root if currently in backend/):

# Return to the repository root if you are in the backend directory
cd ..

# Run the automated pytest suite:
python3 -m pytest tests/test_health.py -v

Expected output:

tests/test_health.py::test_health_check_api_v1 PASSED
tests/test_health.py::test_health_check_root_alias PASSED
tests/test_health.py::test_root_endpoint PASSED
============================== 3 passed in 0.22s ===============================

---

Team & Attribution

• Kasat Sakshi Dattaprasad (@SakshiKasat18)
• Sumit Das (@Sumit-ai-dev)

---

License

This project is licensed under the MIT License — see the LICENSE file for details.