OpenLogos Lab

OpenLogos Lab converts raw ideas and questions into structured research reports. It is designed as a modular scientific institution: each department owns its prompts, schemas, tools, and report fragments, while the core pipeline keeps every action accountable through append-only logs.

Progress Through Questionable Confidence and Excellent Documentation.

Features

• Founder Console with chat and research modes.
• Live board-room panels with parallel department responses.
• Modular research pipeline with labeled output folders.
• LM Studio model routing and health checks.
• Active, passive, and long-term local memory.
• Deterministic demo mode for trying the UI without models.
• Append-only audit logs and redacted diagnostics.

What Happens To A Request

1. A request enters through the CLI, API, Telegram, email, or a future app adapter.
2. The gateway normalizes it into a ResearchRequest and assigns a session ID.
3. The Research Director decomposes the request into research questions.
4. Modular departments investigate the idea from different angles.
5. A feasibility and accountability gate checks risks, contradictions, and weak assumptions.
6. Publications turns the findings into a Markdown research report.
7. Outputs, evidence, department notes, metadata, and audit logs are written to one session folder.

Local Model Setup

The app reads LM Studio configuration from environment variables. Do not hardcode tokens.

cp .env.example .env

Expected LM Studio defaults:

LM_STUDIO_BASE_URL=http://localhost:1234/v1
SALINAS_MODEL_SMALL=google/gemma-4-e4b
SALINAS_MODEL_TECHNICAL=nvidia/nemotron-3-nano-omni
SALINAS_MODEL_ORCHESTRATOR=qwen3.6-27b-mlx
SALINAS_MODEL_SYNTHESIS=qwen/qwen3.6-35b-a3b
SALINAS_CHAT_MAX_WORKERS=8

Run

pip install -e ".[dev]"
openlogos-lab
openlogos-lab run "ideas for apps around local-first AI memory"

Try the UI without LM Studio:

openlogos-lab --demo

Check your environment:

openlogos-lab doctor
openlogos-lab doctor --live

Running bare openlogos-lab opens the Founder Console. It has two modes:

• Chat mode: a board-room round table with department heads. This does not activate the full research pipeline.
• Research mode: wakes the full facility, writes an audit trail, creates a labeled output folder, and produces a report.

Press Shift+Tab in the Founder Console to cycle between Chat and Research modes. Chat mode shows live board-room panels: selected departments wake up, work in parallel, and update as their LM Studio responses arrive. The Director summarizes after the selected departments finish.

Useful chat commands:

/chat              switch to chat mode
/research          switch to research mode
/research this     send the last chat topic into the full research pipeline
/brief             show the latest Director summary
/departments       show board-room seats and routing hints
/models            show LM Studio model mapping
/health            run model health checks
/transcript        show the current board-room transcript path
/memory            search memory for the latest topic
/memory status     show memory counts
/memory search     search memory
/memory reset      clear Lab memory
/remember <text>   store a long-term memory
/remember this     store the latest Director brief
/clear             clear board-room context and start a new transcript
/quit              leave the facility

Board-room transcripts are saved under outputs/boardroom/.

By default, run opens the thematic Lab TUI directly: a terminal facility board where departments wake up, display high-level work bubbles, and log progress as the report is assembled. Use plain mode for scripts:

openlogos-lab run "ideas for apps around local-first AI memory" --no-tui

One-shot board-room chat:

openlogos-lab chat "Should we turn this into a product studio?"

Memory

The Lab has three local memory layers:

• Active memory: per-session working notes.
• Passive memory: append-only observations from chat and research.
• Long-term memory: curated durable facts, stored locally and searched before chat/research.

Hermes-style self-learning is available as an explicit reflection step:

openlogos-lab memory status
openlogos-lab memory search "local-first AI"
openlogos-lab memory add "The founder prefers modular lab departments."
openlogos-lab memory reflect

If you say “remember ...” in chat mode, the Lab promotes that statement to long-term memory after sanitizing it.

The run creates:

outputs/<session-id>_<slug>/
  request.json
  metadata.json
  report.md
  audit.jsonl
  departments/
    director.md
    opportunity_discovery.md
    scientific_inquiry.md
    product_applications.md
    human_testing.md
    risk_ethics.md
  evidence/
    sources.json
    notes.md

Dreaming Mode

The Dreaming Engine monitors configured sources, selects promising topics, and submits them to the same research pipeline as manual requests.

openlogos-lab dream --once

The first implementation uses simple RSS/URL ingestion and deterministic fallbacks so the system can be tested without network or model access.

Development

make dev
make check

Documentation

• Architecture
• Configuration
• Departments
• Demo Mode