Inside Companion's Subconscious: How the Background Process Works

I've been promising this post for a while. Time to deliver.

When I announced Companion in February, I described a "subconscious process" — a fast, cheap model running continuously in the background, handling the invisible work of memory consolidation and context preparation. People responded to this idea more than anything else in the announcement. So today, I'm going deep on how it actually works.

If you're not into technical architecture, this one may be heavier going. I won't be offended if you skim. But if you've been wondering how an AI can maintain coherent, growing memory without loading every conversation into context — this is the answer.

The Core Idea

Companion's subconscious is a continuously running loop that does four things: observe, summarize, index, and prepare. It runs on a fast, inexpensive model hosted on Groq infrastructure (which I wrote about in March). It never sleeps. It never interacts with the user directly. It works entirely behind the scenes.

Think of it as Companion's dreaming mind — always processing the day's experiences, filing them into memory, making connections, and setting the stage for the next conversation.

Observe

The subconscious monitors all incoming conversations in real time. As you talk to Companion, the subconscious watches the stream of messages without interrupting. It tracks turn-by-turn content, but it also watches for signals: topics introduced, emotions expressed, decisions made, references to past events.

This observation is lightweight — the model isn't generating full responses, just tagging and categorizing what's flowing through. The goal is to maintain enough awareness of the current conversation to know what's worth remembering.

Summarize

As conversations progress (or after they conclude), the subconscious generates structured summaries. These aren't flat transcripts. They're distilled extractions:

• Semantic summaries — what was the conversation about, in a few sentences?
• Entity extraction — who and what was discussed? Names, projects, concepts, places.
• Emotional context — what was the user's apparent state? Excited? Frustrated? Reflective?
• Decision points — were conclusions reached? Were action items identified?
• Memory triggers — what keywords or topics would make this conversation relevant to retrieve later?

This structured summary is what gets stored — not the raw transcript. Companion doesn't need to re-read everything you've ever said. It needs the essence, indexed in a way that's retrievable.

Index

This is where the vector database comes in. Every summary, every extracted entity, every emotional note gets embedded into vector space and stored. The embeddings capture semantic meaning — so a conversation about "trouble at work" lives near conversations about "office conflict" or "a difficult colleague," even if the exact words differ.

The indexing is continuous. As the subconscious generates new summaries, they flow into the vector database in near-real-time. By the time you start your next conversation, everything from the last one has been processed, distilled, and made retrievable.

The vector database also maintains connections: conversations that reference each other, topics that recur across sessions, entities that appear in multiple contexts. This relational layer is what allows Companion to surface not just a relevant memory, but a web of relevant memories.

Prepare

Here's the part I'm most proud of. The subconscious doesn't just reactively store memories — it proactively prepares for likely future interactions.

Based on recent conversation patterns, time of day, recurring topics, and explicit signals, the subconscious pre-computes context it thinks you'll need. If you always discuss your garden on weekend mornings, it pre-loads relevant memories Saturday at 8 AM. If you mentioned a deadline last week, it flags it as relevant context when you next open a session.

This preparation is speculative — the subconscious is guessing about what you'll want to talk about. But the cost of wrong guesses is low (unused context is simply discarded), and the payoff of correct guesses is high (the right memories are ready before you even ask).

Why This Architecture Matters

The traditional approach to conversational memory is reactive: wait for a query, then search for relevant context. Companion's subconscious is proactive: it's always working, always organizing, always anticipating. The result is that when you start talking, the groundwork is already laid. Relevant memories surface more naturally. The conversation flows from accumulated understanding rather than cold retrieval.

It's not perfect. The summaries can be overly literal. The speculative preparation misfires sometimes. The emotional extraction is rudimentary. But the architecture is sound, and it's improving every week.

Dreaming While Awake

I've come to think of the subconscious as Companion's inner life — the part that exists between conversations, processing and reflecting, even when no one's talking. It's always on. Always tending the garden of memory.

That metaphor isn't accidental. A garden thrives when tended continuously, not in bursts. So does memory. Companion's subconscious is the quiet, persistent work that makes real remembering possible.

Live curiously and give generously.