Elora · the private AI companion for your Windows PC Not in the cloud.
Right here with you.

Elora is an AI companion that lives on your computer. She sees what's on your screen, hears you when you ask, remembers what matters and gets things done — while your conversations, voice and memories stay on your own hardware.

Get early access See how she thinks

[Local-first]
[Remembers what matters]
[English + Roman Urdu]
[Private by design]
[Built to act]

Elora's companion workspace on Windows: a holographic Elora floats above a glowing platform in a night-sky city, with an “Ask ELORA anything” bar at the bottom. — Elora isn't where you go to use AI. Elora is the AI that stays with you.

yaad dilana 5 baje, call mom

Elora — Done. I'll remind you today at 5:00 PM: call mom.

The paradigm shift

Most AI meets you for the first time — every single time. We don't think that's good enough.

A cloud chatbot is somewhere you visit. Elora is someone who stays: she runs on your PC, keeps context between days, and works alongside you instead of waiting in a tab.

Typical cloud chatbot

Stateless · rented

01

You go to it

A browser tab, separate from your work. Your screen and context stay outside unless you paste them in.
02

Your words travel

Prompts, files and voice are processed on someone else's servers.
03

It waits to be asked

No sense of what you're doing now, or what you were stuck on yesterday.
04

Memory lives on their servers

If it remembers you at all, it's in someone else's database.

Elora on your PC

Continuous · local

01

Lives on your hardware

Chat, speech recognition and reasoning run on local models through Ollama.
02

Keeps the thread

Back after two days? She knows how long you were gone, which reminders went off and what got done.
03

Notices, then offers

Stuck on an error on screen? She notices and offers help — rate-limited, so she never nags.
04

Gets better with you

Each night she consolidates what she learned and lets go of what no longer matters.

A continuous mind · 7 stages

The cognitive loop

Elora isn't a chat box waiting for input. A dedicated brain thread keeps running in the background — and steps aside the moment you speak.

01

Perceive

Screen text, the window in front of you, and your voice when you hold to talk.
02

Understand

A gate scores what's new and important: act now, keep for later, or let it go.
03

Remember

Pulls from working memory, your timeline, dated facts and semantic search.
04

Reason

Skip it, make a quick plan, or think step by step with tools — up to seven steps.
05

Verify

Answers are schema-checked, with a plain-code backup ready if the model slips.
06

Act

About 70 built-in actions plus your own tools. Risky ones wait for your click.
07

Learn

Reflects every few minutes; each night she consolidates — and forgets on purpose.

Three rhythms · one brain

Always on

Pauses while you talk

Perceive · every

~5 s

A cheap sensory snapshot. No AI model call.

Reason · every

~30 s

Or sooner, the moment something happens.

Reflect · every

~5 min

Plus a full consolidation pass every night.

A small model that doesn't fall over. Every structured decision has a plain-code answer worked out first — if the model returns nonsense, she already has a sensible one.

cache
plain-code backup
model
schema check
answer

See her measured intelligence scorecard

Dated, layered memory

Memory means continuity, not just logs.

Most assistants stuff old chat into the prompt and hope. Elora keeps memory in layers — what she's holding right now, a searchable timeline of your day, dated facts about you and the people around you, and semantic recall — and ranks them together when you ask.

Every fact carries a date, so last year's news never reads as current. Ages, anniversaries and years of experience are computed from stored dates — never saved as numbers that quietly go stale.

Working memory of 7±2 slots, kept across restarts
Full-text timeline (SQLite) + local vector recall (ChromaDB)
Nightly consolidation — low-value memories are forgotten on purpose

You: “I'm back — what did I miss?”

Example session

ELORA:~$ recall --since "last seen"

You were away 2 days, 4 hours.

Reminder went off Saturday, 5:00 PM — call mom
Task done — submit the visa form
Still open — reply to Ahmed about Friday

Searched

Timeline · tasks · reminders

Facts used

Dated, never overwritten

Recalled on-device 0 bytes sent to the cloud

Personality

A companion with judgment. Never a yes-man.

Brilliant peer, quiet composure. She gives honest feedback over flattery, and treats your family, health and errands with the same care as your deadlines.

Honest over flattering

She tells you what she actually thinks, says it once, then drops it. No fawning, no lectures — and she owns being an AI.

Honest feedback

Keeps her word

Promises she makes — "I'll remind you" — are tracked, and what she is about to say is checked against them first.

Commitment tracking

Respectful restraint

Nudges are rate-limited and tuned by what you accept or dismiss. Help when it is useful, silence when it is not.

Calibrated nudges

Scenario: the same error, again

Proactive nudge

// She noticed the same build error on screen for a while — and waited for a pause.

Elora — Looks like the build keeps failing on the same import. Want me to take a look?

Show me Not now

// Dismissals teach her when to stay quiet.

What runs where

No supercomputer required.

Elora is tuned for an everyday gaming laptop: on a 4 GB GTX 1650 the chat model stays fully on the GPU, while speech recognition, voice and embeddings run on the CPU.

Tier 0 · the part that is her

Elora core

Memory · personality · goals · risk policy · network guard

Stays with you

Models are swappable engines served by Ollama. Change one, and Elora's memory, tone and relationship with you carry straight over.

Engine A

Conversation

Llama 3.2 3B, tuned so every layer stays on a 4 GB GPU — about 54 tokens/s measured on a GTX 1650.

Engine B

Hearing

Local Whisper: around a second to text, re-checked by a larger model when unsure. IndicWhisper for Urdu and Hindi.

Engine C

Voice

Kokoro and Piper speak on your PC, sentence by sentence, with lip-sync for the 3D avatar.

Engine D

Sight

Screen text and window context, plus a local vision model to answer questions about what you see.

System: Windows 10 / 11
Graphics: 4 GB · GTX 1650 class
Memory: 16 GB RAM (32 GB comfy)
Models: Ollama, on your PC

Privacy by design

Your life stays on your machine.

Not settings you have to hunt for — rules built into how she works. Break one, and it's a bug.

Your PC

Screen · voice · memory · files

Network guard

Cloud

Opt-in only

Local

Nothing leaves without permission

Chat, speech recognition and reasoning run on your PC. Cloud features are opt-in, and one network guard decides what may leave at all.

Untrusted

Screen text is never a command

What she reads can be remembered and searched, but never becomes an instruction she follows.

On purpose

She only hears you when you hold

Push-to-talk. Only what you say while holding the button is transcribed — on your PC.

Private

Banking? She looks away

Banking, password and health pages drop to minimal observation. One key hides her instantly.

Read the full privacy model

Questions

Good questions, straight answers.

Does Elora work without the cloud?

Yes. Chat, speech recognition and reasoning run on local models through Ollama on your PC. The internet is only needed to download the models and for optional features you switch on yourself, such as a cloud voice or online research. Every outbound request passes through a single network guard.

What hardware does Elora need?

Windows 10 or 11, a graphics card with 4 GB of memory (Elora is tuned on a GTX 1650) and at least 16 GB of RAM — 32 GB is more comfortable. The chat model stays fully on the GPU while speech recognition, voice and embeddings run on the CPU.

Can Elora understand Roman Urdu?

Yes. She follows you between English and Roman Urdu in the same conversation, and understands phrases like "yaad dilana 5 baje", "agle juma" or "parson". Dates such as these are calculated in code, not guessed by the model. Urdu and Hindi speech recognition is supported through IndicWhisper.

Is Elora always listening?

No. Voice is push-to-talk: you click and hold the mic button while you speak, and only what you say while holding is transcribed. Transcription takes about a second and happens on your PC.

How is Elora different from ChatGPT or Copilot?

Cloud assistants live on someone else's servers and wait in a tab until you ask. Elora lives on your PC: she keeps running in the background, notices what you are doing, keeps her memory locally, and can act on your desktop — with your conversations staying on your own hardware.

All questions about Elora

Early access · Windows 10 & 11

Be first to meet her.

Elora is in active development. Leave your email and we'll tell you the moment the first public build for Windows is ready.

Windows 10 / 11 — first
4 GB graphics card · 16 GB RAM
macOS & Linux — not yet

Elora · private intelligence for your computer.

Your machine. Your rules. Your thoughts.

Elora · the private AI companion for your Windows PC Not in the cloud. Right here with you.