Alexei — Foundation Release (v0.1) A local AI agent that remembers and is built to learn.
Alexei is a lightweight, persona‑driven AI agent written in C++ and powered by llama.cpp. This early release focuses on the core architecture needed for a future learning AI — a system that can retain memory, adapt over time, and eventually evolve its behavior based on interaction.
Alexei runs entirely on your machine. No cloud. No external APIs. No data leaving your system.
Purpose The goal of this project is to build an AI that can:
retain memory across sessions
adapt behavior based on past interactions
eventually learn from experience without retraining model weights
This release lays the groundwork: persona system, memory structure, modular agent design, and a stable inference loop.
Learning logic (memory extraction, updating, persona evolution) is planned for upcoming versions.
Current Features (v0.1) Local‑only inference using any GGUF model
Persistent memory structure (facts, preferences, persona traits)
Persona engine driven by identity.txt and config JSON
Modular architecture (AI engine, agent core, tools, memory, HTTP)
Simple HTTP interface for sending messages
Clean, stable prompt pipeline
This version focuses on the foundation — the systems that make learning possible.
Project Structure Code src/ Core C++ source code persona/ Persona definition for Alexei config/ Persona configuration JSON memory/ Runtime memory storage (template + placeholders) models/ Place your GGUF model here (not included) third_party/ External dependencies (llama.cpp placeholder) include/ Header-only libraries (e.g., nlohmann/json) Model Weights This repository does not include model weights.
To run Alexei:
Download any compatible GGUF model
Place it in the models/ directory
Update the path in the config if needed
Building This project uses CMake.
Code mkdir build cd build cmake .. cmake --build . Run the executable and Alexei will start listening on the configured port.
Roadmap v0.2 — Learning Begins Memory extraction from conversation
Automatic fact + preference detection
Updating memory.json at runtime
Basic persona adaptation
v0.3 — Behavioral Evolution Trait weighting
Style adaptation
Relevance scoring + memory pruning
v0.4 — Tools & Autonomy Tool execution
Multi‑step reasoning
Expanded HTTP API
Contributing Contributions are welcome. The project is intentionally modular and easy to extend — perfect for experimenting with:
memory systems
agent behavior
persona engines
tool integrations
performance improvements
License MIT License. See LICENSE for details.