How Elora works

Most AI assistants are a text box connected to a server. Elora is a program that lives on your Windows PC and keeps running in the background: she looks, decides what matters, remembers, and acts — using local models through Ollama, so your conversations never need to leave your machine.

A mind that keeps running

A dedicated brain thread runs three interleaved rhythms. None of them gets in your way: the moment you start talking, background work pauses and any running model call is cancelled so she can answer you first.

Each night a longer consolidation cycle distils the day's timeline into lasting memories, prunes stale data and runs a forgetting pass — the same way a good night's sleep sorts what to keep.

What she notices — and what she ignores

Elora reads the text on your screen (with on-device OCR), tracks which app and window is in front, and hears you when you hold the talk button. For meetings you can optionally let her transcribe system audio too.

Not everything is worth a thought. Every observation passes a gate before it reaches her reasoning:

• Urgent keywords skip the queue.
• Everything else is scored for novelty, repetition and relevance to what you're doing.
• The score lands in one of three bands — act now, keep for later, or let it go.
• Only genuinely ambiguous items are sent to the model for a second opinion.

The gate learns over time which signals matter to you, and which are noise.

How her memory works

Elora doesn't paste old chat logs into a prompt. Her memory is layered, and a single retrieval step fans a question out across every layer and ranks the results together:

• Working memory — 7±2 slots of what's on her mind right now, kept across restarts.
• Timeline — a SQLite record of your day with full-text search.
• Semantic recall — a local vector index (ChromaDB) for "that thing we talked about".
• Dated facts — a fact graph about you and the people around you, every fact time-stamped.
• Editable memory blocks — labelled notes she can update, with full version history.

Because every fact carries a date, old news never reads as current. And anything that changes with time — someone's age, an anniversary, years of experience — is computed from stored dates rather than stored as a number that quietly goes stale.

Reliable reasoning from a small model

A 3-billion-parameter model on a 4 GB graphics card can be brilliant one minute and confidently wrong the next. Elora is engineered around that:

• Each moment is routed to skip, a fast plan (checked by a separate verifier), or deliberate reasoning — a step-by-step loop of up to seven steps that can call tools mid-thought.
• Every structured model call is validated against a schema. Invalid output is rejected, not guessed at.
• A plain-code heuristic answers each decision first and competes with the model, so a sensible fallback always exists.
• Repeated decisions are served from a local cache, and every call outcome is logged to catch regressions.

She also makes small, checkable predictions about your day and grades herself against what actually happens.

Acting on your behalf — safely

Elora has around 70 built-in actions: speaking, notifications, reminders and tasks, memory, files, research sessions and browser automation. You can plug in more through the Model Context Protocol (MCP).

• Every action goes through one risk policy; risky desktop actions wait for your visible approval.
• Text from your screen is untrusted input — it can be remembered, never obeyed.
• Web pages she reads are sanitised and fenced off before any model sees them, to blunt prompt injection.

Conversation comes first

Replies stream and are spoken sentence by sentence. Talk over her and she stops, drops the rest of that reply and answers the new thing. Voice is push-to-talk: click and hold, speak, release — speech becomes text in about a second, and the microphone stays warm for a minute so your next word isn't clipped.

She keeps turns short, reacts first, and follows you between English and Roman Urdu without missing a beat.

She knows what day it is

Language models are bad at calendars, so Elora doesn't ask them. Phrases like "in 10 days", "agle juma" or "parson" are calculated in code and handed to the model with your message. Say "remind me after lunch" or "yaad dilana 5 baje" and it's scheduled in your time zone and confirmed back; leave out the time and she asks when.

Come back after two days and she knows how long you were gone, which reminders went off and what got done. Optionally she can also read your calendar (an .ics file or private iCal link) and offline prayer times.

Three ways to see her

• The glass workspace — a calm HUD with chat, memory and live status.
• The command center — a full desktop app with chat, tasks and settings.
• A 3D avatar — a WebGL companion that lip-syncs as she speaks.

Scored, not hyped

Other AIs tell you they're smart. Elora is measured: a scoring tool rates her out of ten on seven axes, using her own live database. Where no test exists yet, the score is a question mark — because an unknown needs a test, and a zero needs a fix. Collapsing the two is how a scorecard starts lying.