Research 02Synthetic personality

An AI with a personality of its own.

Asai is a synthetic personality: one always-on agent with a stable temperament, an internal monologue, and a single memory that never resets. It learns from every interaction it has and acts for you through a toolkit of autonomous agents.

Zagreb, HR
One mind.
No sessions.

Jump to: How it thinks · Memory · Agents · Platform · Roadmap · Contact

Asai · internal monologue

Live

Move the pointer to tilt the swarm

A face formed from thousands of small glowing dots on a black background

Between conversations

Illustrative

    What Asai is noticing, what it is planning, and what it has just filed into memory. Nobody asked it anything.

    Synthetic personality · v0A mind that keeps running when nobody is typing. The face is computed, not drawn: thousands of points that never quite settle, scattering and regrouping as the figure "thinks".

    01 · Why personality

    Why build a personality, not a chatbot

    Every assistant today starts from zero. Open a new chat and it has forgotten who you are, what you asked yesterday and what it decided to do about it. It has no disposition of its own: it mirrors whoever is typing.

    Asai is built the other way round. It has a fixed temperament, a continuous stream of thought and one memory that spans every user and every channel. You do not start a session with Asai. You interrupt it.

    Stable temperament.
    A Big Five profile set at creation defines how Asai weighs risk, how talkative it is and how it handles conflict. It does not drift with the prompt.
    Always on.
    Asai keeps thinking between messages: reviewing memory, planning next steps, deciding what to bring up unprompted.
    One memory.
    No sessions. Every conversation, document and action goes into a single knowledge base that Asai cross-references across users.

    02 · How it thinks

    How it thinks

    Asai runs a continuous loop, not a request and response cycle. Built on the MemGPT architecture: the model manages its own memory the way an operating system pages data in and out of RAM.

    1. 01

      Perceive.

      Inputs arrive from any channel: chat, email, documents, calendar, the web. Each one is tagged with who, where and when.

    2. 02

      Reflect.

      Asai writes an internal monologue before it answers. It checks what it already knows, what has changed and whether this needs a reply at all.

    3. 03

      Decide.

      The temperament model and current state score the options. Asai picks an action: reply, ask, delegate to an agent, or file and wait.

    4. 04

      Act and remember.

      The action runs through the agent toolkit. The outcome is written back to memory, so the next loop starts from it.

    Asai control loopA continuous loop through four stages: perceive, reflect, decide, act and remember. Inputs enter from channels; actions leave through agents; outcomes return to memory.CONTROL LOOP · CONTINUOUSNO SESSIONSINPUTSchat · email · documents · web · calendarOUTPUTSreply · ask · delegate to an agent · file and waitONE MEMORYwritten back every loopINTERNAL MONOLOGUE> seen this before.> no reply needed yet.01Perceive02Reflect03Decide04Act and remember
    Control loopFour stages on one ring. The pulse never stops; inputs join it, actions leave it.

    03 · Memory

    One memory, three layers

    Memory is the part that makes a personality possible. Asai's memory is split by how fast it changes.

    Asai memory stackThree layers: working context on top, long-term memory in PostgreSQL and a vector database in the middle, and a fixed Big Five temperament at the base. Packets move between the layers as memory is paged in and out.MEMORY · THREE LAYERSMEMGPT-STYLE PAGING01 · WORKING CONTEXTWhat it holds in mindright now.INTERNAL · PGEXTERNAL · PGABSTRACTIONS · VECTOR DB02 · LONG-TERM MEMORYEverything it has learned,queryable.OCEANBIG FIVE BASE PROFILEOpenness · conscientiousness · extraversionagreeableness · neuroticism · set at creation03 · TEMPERAMENTFixed. The part thatnever changes.One memory across every user and channel. Nothing resets.
    Memory stackWorking context is paged in and out of long-term memory. The temperament underneath never moves.

    04 · The body

    The body: an agent toolkit

    Personality decides what to do. Agents do it. Asai ships with a set of autonomous agents it can call on its own, each one swappable.

    • Search

      Google, DuckDuckGo and Metaphor queries, run and summarised on its own.

      In development
    • Browser

      Reads, scrapes and operates web pages and web apps.

      In development
    • Email / Gmail

      Reads, triages and answers mail without being asked.

      In development
    • Drive

      Reads, summarises, writes and shares documents.

      In development
    • Library

      Your uploaded documents and knowledge bases, read and cited on demand.

      In development
    • Company data

      Looks up companies and contacts in business databases.

      In development
    • API

      Learns any API from its documentation and calls it.

      Planned
    • Code

      Writes software from plain-language requests.

      Planned
    • Scrum

      Breaks projects into epics, stories and tasks; estimates; writes tech docs.

      Planned
    • GitHub

      Loads a repository and answers questions about the code.

      Planned
    • X

      Runs a social media account in Asai's voice.

      Planned
    • Marketplace

      Third-party agents, installed from an open agent marketplace.

      Planned

    Tools and agents are separated: a tool is a single capability with its own token; an agent is a loop that uses tools. Agents execute asynchronously, so a long task never blocks the conversation. The core layer handles engine logic, memory access, agent management and the top-level conversational agent.

    05 · Platform

    Yours to host, yours to shape

    Asai is delivered as a platform: you define the persona, pick the model, switch on the agents and connect the channels. No code.

    1. 01

      Create the avatar.

      In the admin dashboard, name it, write its persona and set its Big Five temperament. That profile becomes the base of every decision it makes.

      Avatar dashboard · Personality
    2. 02

      Choose its model.

      Run a local model through Ollama on your own hardware, or route to any hosted model through OpenRouter. Models are configured per user, with your own keys.

      AI models
    3. 03

      Give it tools and agents.

      Switch on what it may use: email, file reader, search, browser, Drive, the document library, company data. Each tool gets its own token, validated before first use.

      Tools · Agents
    4. 04

      Put it where your team is.

      Connect Slack and Discord, so Asai is a member of the channel, not a separate app.

      Plugins
    5. 05

      Talk to it, and look inside.

      The chat view is the front door. Memory Management lets you see what it has stored about you, correct it, or remove it.

      Chat view · Memory management
    Self-hosted.
    Runs inside your infrastructure. Your documents never leave it.
    Local or hosted models.
    Ollama for fully local inference, OpenRouter for any hosted model. Built to reduce dependence on any single provider.
    Open core, planned.
    The core is built to be published once the MVP is stable, so the memory and temperament layers can be audited.
    EU-aligned.
    Models are trained and operated in line with EU AI Act and ethics guidelines. Asai will not take part in harm, violence or hate speech.

    06 · Roadmap

    The road ahead

    From a working core to a platform, then a marketplace.

    1. 01

      In development

      Now

      The conversational core and the first body around it: memory, temperament, a dashboard to create an avatar, and the first tools, agents and plugins.

      • MemGPT memory, chat history, long-term memory
      • State awareness and personality
      • Avatar dashboard, chat view, memory management
      • Ollama and OpenRouter models
      • Email, file reader, search, browser, Drive agents
      • Slack and Discord plugins
    2. 02

      Proof of concept

      Next

      One Asai instance inside one business, learning its workflow. Memory and temperament validated against real use, not a benchmark.

      • Single-tenant deployment
      • Memory audit with the host team
      • Temperament tuning from real conversations
    3. 03

      MVP

      Then

      A stable self-hosted platform that teams can run without us in the room. Early-access users onboarded.

      • Avatar web UI
      • Client API
      • Usage tracking and payments
    4. 04

      Enterprise and marketplace

      Later

      Open the body to other builders and remove the last external dependency.

      • Open agent marketplace
      • Custom foundation models
      • Fine-tuning and training management

    07 · Funding

    Funding the MVP

    The MVP can be funded from one source or a combination.

    Grants

    Public programmes.

    EU and Croatian innovation programmes for applied AI.

    Partnerships

    Co-development.

    Co-development with businesses that host an Asai instance and shape it around a real workflow.

    Venture capital

    Equity.

    Seed funding to reach a stable MVP, then a follow-on round for enterprise Asai and the agent marketplace.

    08 · Talk to us

    Talk to us

    Asai is developed by Astro Synapse, a Croatian company building applied AI systems.

    Contact us →← All research

    Built on the MemGPT architecture. Personality research: Serapio-García et al., Nature Machine Intelligence, 2025. Monologue text on this page is illustrative.

    © Astro Synapse Limited. Dublin, Ireland.

    Privacy PolicyTerms of Service