DeepSeek Memory

DeepSeek Memory is a Python-based interactive terminal application that augments a local Large Language Model (LLM) with persistent memory. It leverages Ollama to run a local LLM (e.g., DeepSeek-R1:1.5b or any other model you choose) and maintains conversation history across sessions. This enables personalized, context-aware interactions and allows the AI to recall details like the user's name and prior context over time.

---

Features

• Persistent Memory:
  Conversation history is saved to disk (in dnc_memory.pt), ensuring that user details and context are preserved across sessions.

• Dynamic Context Handling:
  Uses embeddings and usage-based memory management to store and retrieve the most relevant conversation entries.

• Customizable LLM Model:
  Easily switch the underlying LLM by modifying the MODEL_NAME variable in the code.

• Special Query Handling:
  Recognizes and responds to queries like “What’s my name?” or “What’s your name?” using built-in heuristics.

• Interactive Terminal Interface:
  Engage in natural, real-time conversation with the AI directly via the command line.

---

Setup

1. Clone or Download the Repository

git clone https://github.com/yourusername/deepseek-memory.git
cd deepseek-memory

2. Ensure Your Environment Is Set Up

Python:

Install Python 3.8 or higher.

Ollama & Model:

Install Ollama and verify your chosen model works by running:

ollama run deepseek-r1:1.5b

To change the model used by the application, edit the MODEL_NAME variable in the code.

Dependencies:

Install the required Python packages:

pip install torch torchvision torchaudio numpy regex

3. Configure the Model

In the source code (main.py), adjust these global variables as needed:

• MODEL_NAME: Set to the name of the model (default: "deepseek-r1:1.5b").
• BOT_NAME: The bot’s name (default: "Jarvis").

---

Usage

Run the Program

Execute the script from your terminal:

python main.py

Interaction Flow

Startup:

If no user name is stored, the bot (Jarvis) will prompt:

"My name is Jarvis. What's your name?"

The user's response is processed to extract and store only the name (e.g., "Rasim").

Conversation:

• Ask any questions in natural language.
• Special queries like “What’s my name?” or “What’s your name?” are handled by the system.
• For general queries, the recent conversation history is used as context to generate concise, direct answers.

Persistent Memory:

After each interaction, the conversation history is saved to disk (in dnc_memory.pt), ensuring context is preserved across sessions.

---

How It Works

1. Memory Management

Input Conversion:

User inputs are converted into embeddings via the local LLM (using Ollama). If the LLM call fails, a fallback deterministic embedding (using SHA‑256) is used.

Storage:

A fixed-size memory stores these embeddings along with raw conversation text.

Usage Vector:

A usage vector decides which memory slot to update when new entries are added.

2. Special Query Handling

User Name Retrieval:

Queries like “What’s my name?” trigger a scan of the conversation history to retrieve the stored user name.

Bot Name Response:

The bot’s name is hard-coded (default: Jarvis) and is used when the query “What’s your name?” is asked.

3. Response Generation

For general queries, the conversation history (last five entries) is aggregated and sent to the LLM as context, which then generates a direct, concise answer.

---

Troubleshooting

PSReadline Warning

On Windows, you may see a PSReadline warning. This can be safely ignored.

Ollama/Model Errors

If the local LLM call fails:

• Ensure Ollama is installed and the model runs correctly:

  ollama run deepseek-r1:1.5b

• Verify your model configuration and that the MODEL_NAME variable is set correctly.

Memory Persistence Issues

Ensure that the directory where dnc_memory.pt is saved is writable.

---

License

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