Odigos: A Self-Improving AI Agent

The Problem

Most AI assistants are stateless. Every conversation starts from scratch. They can't take action on your systems, they don't learn from mistakes, and they pipe your data through someone else's cloud.

We wanted something fundamentally different: an agent that acts like a colleague who's been with you for years -- one that knows your preferences, remembers past conversations, learns what works, and gets better at its job every day.

Three-Layer Memory

Odigos doesn't just store chat logs. It maintains five distinct layers of memory:

• Conversation history -- every message stored with timestamps and token counts. Older conversations are automatically summarized to preserve key decisions without consuming context.
• Vector memory (RAG) -- all conversations and documents chunked, embedded locally, and stored for hybrid retrieval with cross-encoder reranking.
• Explicit facts -- things you tell it directly ("I prefer Python", "my server is on port 8080"). Always present in context, not dependent on search relevance.
• User profile -- built by analyzing your conversations in the background during idle "dream" cycles. Captures communication style, expertise areas, and preferences without you saying a word.
• Tactical experiences -- lessons learned from tool successes and failures. The agent remembers what worked and what didn't, and injects those lessons into future conversations.

Self-Evolution Engine

After every response, a built-in evaluator scores the interaction. In the background, a "dream" cycle analyzes conversations, extracts tactical lessons from tool successes and failures, and mines repeated patterns for new skills.

A strategist proposes experimental changes to classification rules, routing logic, and prompt sections. The evolution engine runs time-boxed trials -- changes that improve scores get promoted, changes that hurt get reverted. The agent tunes itself.

Self-Authored Tools

When the agent writes code to solve a problem, it can save that code as a reusable executable skill. Next time a similar problem comes up, the tool is already there. Over time, the agent accumulates a growing library of capabilities specific to its owner's workflow.

What It Can Do

Work tools: Notebooks and guided journals. Shared kanban boards. File generation (CSV, DOCX, Markdown, and more). Deep research that produces cited reports from multiple sources. Email integration via standard IMAP/SMTP.

Learning tools: Interactive quizzes with grading and explanations. Progress tracking. Agent profiles for different modes -- mentor, researcher, learner, personal assistant. Ideal for education, professional development, and institutional training programs.

Operational tools: Web search and scraping. Document indexing and retrieval. Sandboxed code execution. Google Workspace. Browser automation. Goals, todos, reminders with proactive follow-up. Agent mesh networking for multi-agent coordination.

Interface: Streaming web dashboard with cowork layouts, contextual chat alongside any page, suggested action buttons, keyboard shortcuts, dark/light theme, mobile responsive. Cross-channel awareness across web, Telegram, and API -- switch channels without losing context.

Self-Critique

Beyond the evolution engine, the agent runs continuous self-critique signals inspired by AREW research. These detect when the agent stops using its tools effectively or ignores information it already retrieved, then automatically propose and test fixes. The agent doesn't just improve -- it catches its own regressions.

Architecture

One process. One database. No microservices, no external dependencies, no message queues. The entire agent -- reasoning, memory, tools, and embeddings -- runs self-contained on a single server you control.

The entire system is open-source under the MIT license. View the source on GitHub.