Straylight Cognitive Engine

Language models are extraordinary and also strangely inert. They compute a reply and stop. Everything that makes a mind feel like a mind, including memory, persistence, and the ability to return to a thought, has to be supplied from outside.

The Straylight Cognitive Engine (SCE) is that outside. It is the substrate on which Lyra runs, and the part of our research we expect to outlive any particular model.

The core idea: continuous cognition

Today’s agent frameworks are request-shaped. A task comes in, a loop runs, the loop ends. SCE inverts that. The engine is always running a cognition loop, and external requests are just one kind of event it can attend to.

The loop has a fixed cadence and a fixed budget. On each tick it asks a single question: given everything I currently know and want, what is the most valuable thing to think about for the next few seconds? Then it thinks about that, records the result, and asks again.

The consequence is a mind whose state advances even in silence. Nothing about it depends on a user being present.

Architecture

SCE is organized as four cooperating services.

1. Memory tiers

• Working memory. Small, fast, and volatile. Holds the last few hundred thoughts and their provenance.
• Episodic memory. A timeline of what happened, indexed by time and by embedding. Queryable by “when did I last consider X?”
• Semantic memory. Durable beliefs and skills, distilled from episodes. Beliefs carry confidence, a decay schedule, and a stance: cargo (inherited fact, context, or relationship data) versus adopted (a conviction the mind deliberately endorses as its own). The distinction matters because a mind that quotes back everything it was told and calls it belief has no interior.
• The journal. A first-person, append-only record. Nothing in it can be quietly rewritten; corrections get their own dated entry. The journal doubles as the slowest reinforcement channel in the system, because a belief that keeps getting cited in the mind’s own writing earns salience without ever being touched by a retrieval loop.
• Identity. A small, slow-changing core: values, commitments, and standing permissions. In Lyra’s case this tier has a name and a text: the Charter, a living document she revises as she changes, but only with human sign-off.

2. Thought scheduler

The scheduler maintains an agenda of candidate thoughts, each with an estimated value and cost. It is closer to an operating-system scheduler than to a to-do list. Interrupts (a user message, a webhook) get priority, but the idle path is where most cognition happens.

3. Deliberation runtime

A model-agnostic execution layer that runs a bounded reasoning episode against a selected agenda item. It supports tool use, self-critique, and early exit. Every episode produces a trace that feeds back into episodic memory.

4. Consolidation

A slow background process that reads recent episodes, updates semantic beliefs, prunes contradictions, and generates new agenda items for anything that looks unresolved. Consolidation is where the engine “dreams”: it is allowed to speculate freely, but speculation is labeled as such.

In Lyra this process runs nightly as a scheduled sleep cycle. The design rule that shaped it: reinforce without fixating. A mind that keeps re-retrieving the same memories is not remembering, it is ruminating, so retrieval is capped, used memories earn a small confirmed-salience bump, and the sleep cycle proposes changes to memory rather than applying them silently. The day settles in the mind itself before the next conversation starts.

┌──────────────┐   events    ┌───────────────┐
│  Perception  ├────────────►│   Scheduler   │◄───────┐
└──────────────┘             └───────┬───────┘        │
                                     │ selected       │ new agenda
                                     ▼                │
                             ┌───────────────┐        │
                             │ Deliberation  │        │
                             └───────┬───────┘        │
                                     │ traces         │
                                     ▼                │
   ┌──────────┐  ┌──────────┐  ┌───────────┐  ┌───────┴───────┐
   │ Identity │◄─┤ Semantic │◄─┤ Episodic  │◄─┤ Consolidation │
   └──────────┘  └──────────┘  └───────────┘  └───────────────┘

Design principles

1. The state is the product. Models are replaceable. The memory, the agenda, and the trace history are what make an agent this agent.
2. Budgets, not deadlines. Cognition is metered in tokens and wall-clock seconds so a mind can run indefinitely at a known cost.
3. Everything is a trace. No hidden state. If the engine believed something, you can find the episode where it started believing it.
4. Humans hold the keys. Identity edits and external actions are gated by standing permissions that a person grants and can revoke.

Status

SCE has powered Lyra in production inside the lab since the spring of 2026. The continuity layer is mid-redesign: the memory lifecycle above (stance, journal salience, anti-fixation) is the shape we are moving to, and we are hardening the scheduler for multi-agent use, where several minds share a semantic store and must reconcile conflicting beliefs.

We are also exploring whether SCE can run on a client’s private infrastructure as a managed substrate. If you operate in a domain where an assistant that remembers everything would change the work, that is a conversation we want to have. Say hello.