🐻 Memory Bear: Cognitive Edge AI on Legacy Hardware

"Intelligence is not about how much you know, but how effectively you filter what you forget."

📖 Executive Summary

Memory Bear is a local, cognitive AI agent engineered to run efficiently on highly constrained legacy hardware (Intel-based MacBook Air 2017).

Standard Large Language Models (LLMs) suffer from "context amnesia" or inefficient RAG retrievals that flood the context window with noise. This project implements the Memory Bear Architecture [1], utilizing a biologically inspired Forgetting Curve to dynamically manage context. This allows a small SLM (Phi-3-Mini) to maintain coherent, long-term conversations without exceeding the 8GB RAM limit of legacy devices.

---

🏗️ System Architecture

The agent moves beyond simple "Golden Retriever" RAG by implementing a Cognitive Loop that mimics human memory consolidation:

graph TD
    A["User Input"] --> B{"Hippocampus Retrieval"}
    B -->|"Vector Similarity"| C["Episodic Memory"]
    B -->|"Graph Traversal"| D["Semantic Cortex"]
    C --> E["Ebbinghaus Filter"]
    E -->|"Decay Check: R > 0.35?"| F["Filtered Context"]
    F --> G["LLM Orchestration"]
    D --> G
    G --> H["Generative Inference Phi-3"]
    H --> I["Output to User"]
    H --> J["Consolidation Save & Reinforce"]

Loading

Core Components

• Hippocampus (Episodic Memory): Uses ChromaDB to store temporal interaction logs.
• Cortex (Semantic Memory): Uses NetworkX to build a Knowledge Graph of static facts.
• Cognitive Filter: Applies the Ebbinghaus Decay Formula ($R = e^{-t/S}$) to prioritize "fresh" or "rehearsed" memories, discarding noise.

---

🛠️ Engineering Case Study: Optimization for Legacy Hardware

Developing modern AI on a 2017 MacBook Air (Dual-Core i5, 8GB RAM, No GPU) presented specific engineering constraints.

Constraint	The Challenge	Engineering Solution
Compute	No GPU Acceleration (Metal/CUDA unavailable)	Utilized llama.cpp for CPU-optimized inference with AVX instruction sets.
Memory	8GB Total RAM (Shared with OS)	Implemented Quantization (q4_k_m) to compress the 3.8B parameter model to ~2.4GB RAM usage.
Latency	Slow Inference (2-3 tokens/sec)	Engineered a "Thinking" UI loop to manage user expectations and reduce perceived latency.
Dependencies	macOS 12 (Monterey) Compatibility	Resolved critical ABI conflicts between NumPy 2.x and PyTorch by pinning stable legacy versions.

---

📊 Exploratory Data Analysis (EDA)

To validate the biological plausibility of the architecture, I conducted a simulation of the memory decay algorithms before deployment.

• Objective: Determine the optimal "Forget Threshold" to maximize context relevance.
• Result: Setting the retention threshold to 0.35 successfully filters 90% of "Noise" (S=1) within 1 hour, while retaining "Core Memories" (S=20) for >24 hours.
• Full Analysis: See notebooks/memory_dynamics_eda.ipynb

---

🚀 Installation & Usage

Prerequisites:

• Anaconda / Miniconda
• macOS 12+ (Optimized for Intel Macs), Linux, or Windows

1. Environment Setup

We utilize a dedicated environment to ensure stability on older CPU architectures.

# Create environment
conda create -n memory_bear_legacy python=3.10 -y
conda activate memory_bear_legacy

# Install dependencies (pinned for stability)
pip install "numpy<2.0"
pip install -r requirements.txt

2. Run the Agent

python main.py

Note: The first run will automatically download the quantized Phi-3-Mini model (~2.4GB). Please allow 5-10 minutes depending on your internet connection.

🔮 Roadmap & Future Work

Phase 1 (Completed): Core cognitive architecture on CPU.

Phase 2 (In Progress): Multi-modal integration (Image inputs via Llava).

Phase 3 (Planned): Implementation of "Dreaming" (Offline consolidation) where the agent summarizes daily logs into long-term semantic nodes during system idle time.

📚 References

Memory Bear AI (2025). A Breakthrough from Memory to Cognition. Shanghai Shanmo Shongyang Technology.

Vaswani et al. (2017). Attention Is All You Need. Google Brain.

Engineered by Lim Wen Gio