BNOS AI Companion

🌍 Language: English | 中文

A fully local AI digital companion — a single AI organism, not a chatbot, not a multi-agent platform.

---

Overview

BNOS AI Companion is a fully local, single-entity AI companion. It is built on the BNOS orchestration engine and integrates cognition, memory, emotion evolution, tool use and knowledge management into one AI organism — an entity with its own memory, its own personality, and its own autonomous behavior.

The companion is structured like an organism: each "organ" is an independent node running in its own OS process and virtual environment. The BNOS engine is the nervous system that schedules them, and all organs communicate through a file-based JSON protocol.

This project demonstrates the orchestration capability of BNOS applied to a complete, local-first AI product.

---

Core Design Philosophy

Principle	Implementation	Commitment
AI is an independent entity	Own memory, own personality, own thoughts	Not merely a user tool
Fully local	All data and models stored & run locally	User owns all privacy
Unlimited growth	Node-level orchestration, capabilities keep expanding	No functional ceiling
Process-level isolation	Each node: independent process + independent venv	Crashes don't cascade

The Organism Metaphor

AI Part	Corresponding Component	Responsibility
🧠 Brain	aaa_cognition + memos.py	Cognition loop, memory read/write, emotion evolution
👤 Face	live2d_face + tts	Live2D expressions, TTS speech synthesis
🖐️ Hands	node_dsh (DeepSeek Harness)	Tool calls: on-demand tool assignment, agent execution, MCP extensions
🐚 Hippocampus	logseq_writer	Knowledge graph, long-term document archiving
⚡ Nervous system	BNOS engine	DAG orchestration, process scheduling, file protocol

---

Architecture

Star Topology: One Cognition Hub, Many Inputs

All input sources converge on the AAA cognition hub (the single memory entry point), which routes through one output port with data_type routing:

• prompt → LLM
• tool_call → DSH (DeepSeek Harness tools)
• reply → Live2D face
• knowledge → Logseq writer

ASR / GUI / env input ──→  aaa_cognition ──→  llm_infer ──→  aaa_cognition ──→  live2d_face (display)
    (3-phase prompt,         ↑                (parse,          └──→ tts (speech)
     MemOS semantic           │                write DB,
     retrieval,               └── memos        index rebuild)
     identity_key)               (built-in)

                      node_dsh (DSH tool execution)
                      logseq_writer (knowledge persistence)

Tech Stack

Layer	Technology	Notes
Orchestration	BNOS (Python/PySide6)	IDE for development, lightweight engine at runtime
Memory	Python + SQLite + MemOS (numpy)	AAA cognition loop + vector semantic retrieval
LLM inference	llama.cpp + cloud API	Dual backends, one-click switch
Live2D rendering	PixiJS + Cubism SDK 4.x	Extracted from My-Neuro
TTS	edge-tts + MOSS-TTS-Local	Online + local dual channel
Tool execution	node_dsh (DeepSeek Harness)	Agent loop + tool pool + MCP extensions
Knowledge graph	Logseq	Markdown + bidirectional links
GUI client	PySide6	Lightweight, non-web
Communication	File JSON	stdin/stdout + output.json

---

Node Matrix

#	Node	Language	Role	Status
1	node_python_aaa_cognition	Python	Data hub: 3-phase prompt + MemOS + tag parsing	🟢 Core chain complete
2	node_python_llm_infer	Python	LLM inference: cloud API + local GGUF	🟢 Cloud API integrated
3	node_js_live2d_face	JS	Character display: Live2D Cubism 4.x	🟢 Core logic complete
4	node_python_tts	Python	Speech synthesis: edge-tts + MOSS local	🟢 Basic usable
5	node_python_asr_input	Python	Speech recognition: Silero VAD + SenseVoice	🔴 Design finalized
6	node_python_env_input	Python	Environment sensing: CPU / memory / time	🔴 Skeleton exists
7	node_python_logseq_writer	Python	Knowledge archiving: Markdown + backlinks	🟡 Generates .md, not writing to disk
8	node_dsh	Python + TS	Tool execution: DSH agent loop + tool pool + MCP	🟢 Integrated
9	node_python_vlm	Python	Multimodal vision: screen / camera / image	🔴 To be created

Key Subsystems

Subsystem	Location	Core Capability
MemOS semantic retrieval	aaa_cognition/memos.py	SentenceTransformer encoding + numpy cosine similarity + decay
3-phase prompt	aaa_cognition/prompt.py	Thin prompt → LLM decides retrieval → second interaction with results
identity_key isolation	Full pipeline	Multi-user data isolation, vector space partitioned per user
turn_taking filter	AAA internals	Rule filtering + observation buffer + hysteresis loop
Personality evolution	AAA planned	4-dim personality vector (warm/lively/direct/curious) + passive feedback

---

Key Features

• AAA cognition loop — a deterministic cognition cycle: perceive → retrieve → reason → act → remember, with a 3-phase prompt design.
• Memory evolution — SQLite + MemOS vector retrieval with decay, letting the companion's memory grow and fade like a real one.
• Cross-vendor LLM experiments — DeepSeek / Qwen consistency tests and long-term memory evolution studies (see tests/, scripts/aaa_compare/).
• Traceable agent behavior — every node writes structured JSON; the full chain GUI → LLM → face is auditable.
• Fully local GUI — PySide6 dashboard (node status / CPU / memory) plus a chat page and knowledge-base panel.
• Research subproject schemanet/ — an independent study on gradient-free structural learning (spiking networks with Hebbian/STDP rules). See its own README.

---

Quick Start

# 1. Clone
git clone https://github.com/LiuStar656/BNOS-AI.git
cd BNOS-AI

# 2. Start all nodes
run.bat        # Windows
./run.sh       # Linux / macOS

# 3. The PySide6 GUI opens automatically — chat with the AI in the "Chat" page

API keys are read from environment variables only (DEEPSEEK_API_KEY, QWEN_API_KEY). No keys are hardcoded.

---

Project Structure

BNOS_AI_project/
├── bnos_runtime/            # BNOS runtime engine (engine, pipeline loader, runner)
├── nodes/                   # Core nodes (each: independent process + venv)
│   ├── node_python_aaa_cognition/   # 🧠 cognition hub (memos.py / prompt.py / db.py)
│   ├── node_python_llm_infer/       # ⚡ LLM inference
│   ├── node_js_live2d_face/         # 👤 Live2D face
│   ├── node_python_tts/             # 🔊 speech synthesis
│   ├── node_python_asr_input/       # 👂 speech recognition (planned)
│   ├── node_python_env_input/       # 🌡️ environment sensing
│   ├── node_python_logseq_writer/   # 📝 knowledge archiving
│   └── node_dsh/                    # 🖐️ tool execution (DeepSeek Harness, upstream in harness/)
├── gui/                     # PySide6 client (dashboard / chat / knowledge / settings)
├── docs/                    # design docs & architecture docs
├── schemanet/               # research subproject (gradient-free structural learning)
├── tests/                   # consistency / evolution tests
├── scripts/                 # experiment scripts (aaa_compare, ...)
├── pipeline.json            # core pipeline declaration
└── run.bat / run.sh         # launchers

---

Testing & Experiments

• tests/llm_consistency_test.py — cross-round consistency of the same persona.
• tests/self_evolution_test.py — memory/emotion evolution over many rounds.
• tests/message_pool/ — interest-gated multi-source message platform experiments.
• scripts/aaa_compare/ — cognition design comparisons (e1e2 / e4 / e6).
• schemanet/ — see its report in schemanet/docs/.

---

Documentation

Doc	Status	Description
BNOS-AI 伴侣开发方案	Design master	Initial architecture & core chain (Chinese)
节点开发规范	Current spec	Node development standard (Chinese)
node_config_json 开发规范	Current spec	node_config.json schema (Chinese)
docs/design/	PLAN	Feature design docs (3D character, personality seed, event-driven behavior, ...)

---

Known Limitations

• TTS currently online-only (edge-tts); local engines planned.
• ASR, VLM and environment input nodes are designed but not yet implemented.
• logseq_writer generates Markdown but does not write to the Logseq directory yet.
• Local GGUF inference is wired but cloud API is the primary tested path.

---

License

MIT © 2026 Ahdong&Shouey