🧠 Ulisse Memo

A persistent memory infrastructure for LLMs.
Installable. Customizable. Private.

Turn any LLM into an entity with continuous memory.
Not just a chatbot — a full cognitive architecture.

---

✨ What is Ulisse?

Ulisse is a system that allows a language model to remember what you told it, even across different sessions. It's an architecture that turns a generic LLM into an entity with continuous memory, capable of updating its own knowledge as it talks to you.

💻 Local Models

Ollama, LM Studio, or any
OpenAI-compatible server

🔑 Remote APIs

OpenAI, DeepSeek, Claude,
or any provider with an API key

---

🏗️ Architecture — Three Memory Layers

Ulisse is built on a three-layer cognitive architecture. The user doesn't need to configure any of this — it works out of the box.

graph TD
    subgraph "💬 Layer 1 — Current Conversation"
        A[User Message] --> B[AI Response]
        B --> A
    end

    subgraph "🔍 Layer 2 — Short-Term Memory · STM"
        C[(ChromaDB\nVector Store)]
        D[Semantic Search\n& Retrieval]
        C --> D
    end

    subgraph "📚 Layer 3 — Long-Term Memory · LTM"
        E[Semantic Wiki\nMarkdown-based]
        F[AI-Managed\nKnowledge Graph]
        E --> F
    end

    A -->|RAG Query| D
    D -->|Relevant Context| B
    B -->|Synthesize & Store| E
    A -->|Direct Lookup| E

    style A fill:#0d1b2a,stroke:#00d4ff,color:#fff
    style B fill:#0d1b2a,stroke:#00d4ff,color:#fff
    style C fill:#1a0a2e,stroke:#a855f7,color:#fff
    style D fill:#1a0a2e,stroke:#a855f7,color:#fff
    style E fill:#1a1a00,stroke:#ffd700,color:#fff
    style F fill:#1a1a00,stroke:#ffd700,color:#fff

Loading

Layer	Name	Technology	Purpose
1️⃣	Current Conversation	In-context window	Immediate chat context
2️⃣	STM (RAG)	ChromaDB vectors	Context-aware retrieval from past sessions & docs
3️⃣	LTM (Wiki)	Markdown semantic wiki	AI-managed persistent knowledge, projects & facts

---

⚡ Capabilities & Toolkits

Ulisse is not just a chatbot — it is an agentic system equipped with a specialized sub-agent and native tools to interact with the real world.

🛠️ Native Tools Directly built into the core for maximum reliability Tool Description 📂 Workspace Reader List files and folders in your project 📄 File Reader Open and read any file within the workspace 📝 Wiki Manager Read, write, and organize semantic long-term memory

🤖 Agno Agent (Sub-Agent) A powerful autonomous engine for complex tasks Tool Description 🌐 Web Search Browse the internet for real-time information 🐍 Python Runner Execute Python scripts for complex logic 💻 Shell Access Execute terminal commands in the workspace 📊 CSV Analysis Query large datasets via DuckDB SQL 🕸️ Browser Advanced web scraping (optional)

---

🚀 Installation

Prerequisites

• Python 3.11+
• An LLM API key OR Ollama/LM Studio running locally

Note

First-time setup: On the first launch, the system will automatically download all-MiniLM-L6-v2 (~80 MB) for RAG embeddings and ChromaDB dependencies. These are cached locally in ./hf_cache and ./vectordb. Subsequent startups will be faster.

Quick Start

# 1. Clone the repository
git clone https://github.com/FridaAlma/UlisseAI.git
cd UlisseAI

# 2. Install dependencies
pip install -r requirements.txt

# 3. Configure environment
cp .env.example .env

Edit .env with your settings:

DEEPSEEK_API_KEY=your_key_here
DEEPSEEK_BASE_URL=https://api.deepseek.com
CORPUS_PATH=./corpus
VAULT_PATH=./vault
VECTORDB_PATH=./vectordb

# 4. Launch
python webapp/backend/app.py

Then open http://localhost:5000 🎉

---

🌐 Choosing the AI Model

In the chat interface, next to the Send button you'll find a 🌐 network button that opens a provider selector:

Option	Icon	Description
LLM Locale	💻	Connects to a local model (Ollama, LM Studio, etc.). Uses DEEPSEEK_BASE_URL / DEEPSEEK_API_KEY from .env
API Key	🔑	Enter any provider's Base URL, API Key, and model name directly from the UI. Saved in localStorage
Ulisse Memo v1	🧠	[EXPERIMENTAL] Local MSA-4B model with persistent memory bank. (Under active development)

Your choice is persisted in the browser across reloads.

⚙️ Advanced: Changing the Model Manually

Option A — Local / Default model

Edit .env:

DEEPSEEK_API_KEY=your_key_here
DEEPSEEK_BASE_URL=https://api.deepseek.com   # or http://localhost:11434/v1 for Ollama

Then in webapp/backend/app.py, find and change:

chat_model = "deepseek-chat"  # → "gpt-4o", "llama3", "qwen2.5:7b", etc.

Option B — Fully custom provider in code

In webapp/backend/app.py, locate the provider routing block (~line 408):

# === Provider routing ===
provider = data.get("provider", "local")

You can add new branches here to support additional providers at the server level.

---

🧬 System Prompt

Ulisse operates using a specialized system prompt located in corpus/system_prompt.md.

Tip

Customizing the System Prompt: You can personalize Ulisse's behavior, but proceed with care:

• ✅ Identity and Role — Freely modify to redefine who Ulisse is
• ⚠️ Personality and Tone — Adjustable, but heavy changes may increase hallucinations. Adaptability and Irony are safe to customize
• 🚫 Technical Instructions — Do not modify. Memory architecture (STM/LTM) and Wiki management depend on specific instructions

---

🛠️ Tech Stack

Component	Technology
Runtime	Python 3.11+
Vector DB	ChromaDB
LLM Provider	DeepSeek / OpenAI-compatible
Long-Term Memory	Semantic Wiki (Markdown)
Knowledge Graph	Obsidian-compatible
Backend	Flask
Frontend	Vanilla JS (single-page application)

---

📊 Project Status

██████████████████░░░░░ ~75%

Active Development — Functional and in testing phase. Note: The "Ulisse Memo v1" (MSA-4B) mode is currently in Beta/Development phase. Expect high VRAM usage and potential instabilities. Contributions, bug reports, and ideas are welcome!

---

📜 License

This project is licensed under the Apache License 2.0.

---

_{Built with 🧠 by FridaAlma
Report Bug · Request Feature}