DepthNet: Platform for 'Living' Autonomous AI Agents with Advanced Memory — One Year Later
Anthropic
OpenAI
DeepSeek
A year after publishing a prototype, the author has evolved it into DepthNet, an open-source platform for autonomous AI agents. The platform now features multiple memory types, RAG, orchestration of multi-agent pipelines, vision, voice, and a layer of internal state. It supports two modes: free agents and predictable orchestrated pipelines, and can be installed via Docker.
DepthNet is an open-source platform (rnr1721/depthnet) that has evolved from a simple prototype into a full-featured system for autonomous AI agents. Key additions include several memory types (notepad, vector memory, journal, skills, ontology, workspace, persons), RAG with configurable pipelines, orchestration of multi-agent teams (planner, executor, critic, validator) managed by deterministic code, and vision and voice capabilities. The agent can now use native tool calls (function calling) in addition to legacy text tags. The platform operates in two modes: free agents (unrestricted chat or continuous loop) and orchestrated pipelines (predictable tasks with roles). Memory is structured, allowing the agent to recall episodic events and maintain an ontology of entities and relationships. RAG is implemented as a configurable pipeline where a small model formulates search queries, and the agent can actively queue retrieval requests. The system also features hysteresis: it switches between a rich RAG context mode for conversation and a rich procedural context mode for long tasks, based on activity detection of instrumental plugins like browser, terminal, code, and Telegram.
- Abbreviations
- RAG = Retrieval-Augmented Generation — генерация с дополнением извлечением
- LLM = Large Language Model — большая языковая модель
- PHP = PHP: Hypertext Preprocessor — язык программирования PHP
- tf-idf = Term Frequency-Inverse Document Frequency — частота термина-обратная частота документа
Source: Habr — хаб ИИ —
original