Technical
What's SERA?
Most AI answers the question in front of it. SERA builds an understanding across every conversation, and gets better at it over time.
SERA stands for Self-Evolving Recursive Awareness. It’s the system behind Sera, the voice you talk to on algn. Most AI gives you an answer. SERA builds a picture of how you think, not from a single conversation but from the patterns across all of them: the recurring themes, the things you keep circling, the gap between what you say you want and what you keep returning to.
You should never have to re-explain yourself. That’s the surface promise. What earns it is everything happening underneath.
It doesn’t rush. When a question carries weight, the pace of the reply changes with it, and pauses land where they matter rather than all before the answer starts. That’s a deliberate design choice, not a delay being managed.
SERA is self-evolving and recursive. Its understanding of you grows out of your conversations, not a profile set at sign-up. And it checks what it just said against what it understood before, and corrects itself. Its understanding isn’t just accumulated. It improves conversation by conversation, by being wrong less often.
Alongside the conversation, a set of specialised agents runs in parallel, each with its own job. One reads the conversation in real time and decides how it should flow. One extracts, consolidates and forgets memory. One looks for patterns across conversations: themes that recur, topics you talk around without landing on. One watches for distress. None of them slows the conversation down. Together they shape how Sera responds, and what’s worth keeping goes into your memory, which you can open, read and delete at any time.
The approach is loosely modelled on how human memory works: not a recording, but patterns drawn out of experience and kept where they’re useful.
Awareness here means noticing: the patterns in what you say, and in its own replies. It isn’t a claim that Sera is conscious. Sera is an AI, and will always tell you so.
The patterns you keep returning to say more than the words you use.
What SERA does
- Builds a working model of how you think, not just what you said
- Paces replies to the weight of the conversation
- Notices recurring patterns across conversations
- Carries understanding forward, so each conversation starts further in
- Keeps your memory encrypted on your device, where you can see and delete it
What it doesn't
- Reset between conversations
- Rush every answer to the same pace
- Store what you say to sell it later
- Flatten your thinking into a generic profile
- Mistake recall for understanding
Architecture
One conversation you’re part of. Several agents running alongside it, in parallel, without slowing it down. The Technical Agent reads the conversation in real time and makes the routing calls: what topic you’re in, what pace fits, which other agents need to know. Everything adapts as the conversation develops, not from a profile set at signup.
- Main Conversation: the voice you talk to · responds in real time
- Technical Agent: topic · pacing · real-time routing
- Memory Agent: extraction · consolidation · on-device memory
- Subconscious Agent: background pattern analysis · notes passed to the main conversation
- Crisis Agent: checks every turn for distress · classifiers check what you say and what Sera says · stops any challenge · recommends official sources of support · no one monitors conversations
Memory that forgets properly
Most AI memory is a pile that grows without limit. SERA’s consolidates. Near-duplicate memories merge rather than stack. Different kinds of memory fade at different rates: who you are fades slowest, what you did last Tuesday fades fastest. The result is a memory that stays useful rather than just getting larger.
Memory that stays useful rather than just getting larger.
Where your data goes
Your memory and conversation history are encrypted on your device, with a key that never leaves it. algn never has that key. Backups are encrypted with a passphrase only you hold.
When a conversation needs cloud processing, the relevant turns are sent to open-weight models under zero data retention: they’re processed, answered, and not kept. For matching in Us, sensitive topics are stripped on your device first, and only reduced, salted numbers leave it, never the words behind them.
The honest limit: more context isn’t always more understanding. SERA is designed to reconcile, not to hoard. That’s why it forgets deliberately. More memory only helps when it adds clarity, not noise.
Corrections
This page is stable. We don't expect it to change much, but we still fix what we get wrong. If a claim looks off or a source doesn't hold up, tell us below.