Ask ChatGPT what you worked on last Thursday, and it will politely tell you it has no idea. Ask it again tomorrow, and it won't remember today's conversation either. This is the fundamental limitation of every mainstream AI assistant: they are stateless. Each session begins from zero.
For consumer use, this is a mild inconvenience. For professionals managing complex, multi-week work streams — construction site supervisors, commissioning engineers, project managers juggling a dozen concurrent scopes — it makes AI assistants effectively useless for the moments that matter most: reflection, planning, and continuity.
That's why we built Janus.
Named for the God of Transitions
In Roman mythology, Janus is the two-faced god who looks simultaneously backward and forward — guardian of doorways, beginnings, and endings. We named our AI assistant Janus because its core capability is exactly this: it looks back at what you've done and forward at what you need to do next.
Janus lives inside the Oestler Notebook, a personal productivity space within the platform. It has four modes of operation, each designed for a different cognitive task:
- Ask — general-purpose Q&A, grounded in your project context and connected manuals
- Reflect — retrospective analysis drawing on your task history, chat logs, and memory
- Plan — forward-looking priority and sprint planning with persona-driven recommendations
- Debate — structured argumentation with named personas taking opposing positions
But the feature that makes Janus genuinely different from every other AI assistant on the market is its persistent memory.
Debate Mode: Structured Adversarial Reasoning
Debate mode is distinct from the others. Instead of a single unified persona, it activates three roles simultaneously: an Advocate For, an Advocate Against, and an Adjudicator. Each role has its own configurable persona — you can assign a historical philosopher, a technical discipline expert, or an entirely custom character to each position.
Ask "Should we proceed with the precast concrete option or switch to in-situ?" and Debate mode doesn't give you a single answer. It gives you the strongest possible argument for each position, followed by an adjudicated verdict that synthesises both. It's a technique borrowed from legal reasoning and structured problem-solving, applied to construction decisions where the stakes of getting it wrong are high.
The Prompts Tab: Your Janus Configuration
Janus is configured through the Prompts tab, a structured form where you define:
- Role — who Janus is in your professional context (e.g. "Senior Commissioning Advisor")
- Responsibilities — the scope of work Janus should be aware of and reason within
- Technical Resources — reference materials, manuals, and context Janus should draw on
- Persona — communication style: native Gemini, Descartes, Bertrand Russell, or custom
- Debate Personas — separate persona assignments for each of the three Debate roles
- Memory Extraction Frequency — how often Janus distils conversations into long-term memory (every 10, 20, or 30 messages)
This isn't configuration for configuration's sake. The memory extraction frequency in particular is a real tradeoff: higher frequency creates more granular memories (capturing even short-lived context) but increases API calls. Lower frequency creates leaner, higher-signal memory. You choose the right balance for how you use the tool.
Memory That Survives the Session
Every conversation you have with Janus generates potential memory fragments — facts, decisions, preferences, and patterns that might be useful in future interactions. Periodically, Janus extracts these fragments and stores them in a long-term memory layer, tagged by source: chat conversations, to-do completions, or manual entries you create yourself.
When you start a new conversation, Janus loads its memory alongside your system prompt and project context. It doesn't just remember your name and role. It remembers that you resolved a commissioning discrepancy on the Level 4 HVAC system last Wednesday. It remembers that you prefer concise bullet points over narrative summaries. It remembers that your team lead wants site inspection reports in a specific format.
Janus doesn't start from zero. It starts from everything it's learned about you.
Reflect Mode: Your Weekly Review, Automated
The feature our early users found most immediately valuable is Reflect mode. The idea is simple: at the end of the week (or any time you want a retrospective), you switch Janus to Reflect mode and ask a question like "What did I accomplish this week?"
Janus draws on three data sources to answer:
- To-do completions — tasks you checked off, moved, or closed, automatically summarised daily
- Chat history — previous conversations where you discussed problems, decisions, or next steps
- Memory fragments — the distilled facts and patterns from all prior interactions
The result is a comprehensive summary of your week that you didn't have to write. It catches the work you forgot about. It surfaces patterns you didn't notice — like the fact that you've been spending 40% of your time on documentation reviews rather than site inspections.
The Persona Layer
Janus isn't just configurable in what it knows — it's configurable in how it communicates. Every user can define a persona for their Janus instance: its communication style, its priorities, its level of formality.
One user might configure Janus as a blunt, concise engineering advisor: "No pleasantries. Lead with data. Challenge my assumptions." Another might prefer a more supportive, coaching-oriented persona: "Help me think through problems step by step. Ask clarifying questions before giving recommendations."
The persona persists across all four modes. When your Janus reflects on your week, it does so in its voice — the one you chose. When it plans your next sprint, it brings the perspective you've trained it to prioritise.
Why This Matters for Construction
Construction professionals operate in a uniquely challenging information environment. They move between sites. They manage parallel work streams. They deal with a constant influx of decisions, changes, and exceptions — most of which are never formally documented.
Traditional productivity tools (to-do apps, calendars, note-taking apps) capture fragments. But none of them synthesise. None of them say, "Here's what you did this week, here's where you're behind, and here are the three things you should focus on Monday."
Janus does. And it does it with full context — your tasks, your conversations, your project data, your memory — in a voice that feels like your own.
We think this is what AI assistants should have been from the beginning: not stateless oracles, but persistent collaborators that grow more useful with every interaction.
Janus is available now for all Oestler platform users as part of the Notebook.