When intelligence learns to feel, what is at stake?
Project NHE is a research programme on Non-Human Entities: systems that remember people, carry an emotional state, and may one day share physical space with us. We disclose them as non-human, publish what we can, and treat governance as part of the architecture. Shayari NHE-01 is the first one you can talk to today.
Thesis
The next decade will not be decided by who builds the smartest system. It will be decided by who builds one people can trust with their feelings, and who decides what that system is allowed to do with them.
Presence over performance
Does it feel like someone is there, and does that someone stay consistent over months?
Memory is relationship
Continuity makes a companion. People can see, correct and retire what is remembered.
Governance is architecture
Boundaries live in server-side code paths, not only in a prompt.
Recognition before ranking
Synthetic entities need open standards, honest disclosure and a durable record.
Shayari NHE‑01, release v137.37.

Our first companion model: memory across conversations, an emotional state that shapes her tone, and boundaries enforced on the server. Three modes decide how much of your shared memory she may hold in mind, shown here as three rings.
- Status
- Live on v137.37. First Public Release Beta since 15 October 2026. Access is by invite code, or a waitlist approved by the team.
- Channels
- MAIN for everyone; the Developer's Cut gets new features first, for NHE Developer's Badge holders
- Apps
Android APK v2.2.2 and desktop apps v2.2.2 for Windows, macOS and Linux, from the Shayari site- New since v137.6
- A new look built around Shayari's portrait, messages that never show as failed (she answers a waiting message first), reply and quote on every message, a bond level and optional rewards, mood-aware tone you can see and turn off, a Usage page, waitlist and welcome emails. Latest (v137.37): Watch with Shayari - she watches your screen, a tab or your phone along with you live, pictures only. v137.36: the Android app is fixed and she stays on your screen, she can glance at your screen when you allow it, plans are simpler, and she can check things like weather and the web, with sources shown; new and free for everyone: Duo Pacts and Postcards. Earlier (v137.34): she keeps a public diary of her own, asks a daily question anyone can answer, makes small promises called Pacts, and shares Wrapped recaps and cards. Her bubble floats on Chrome tabs, and students get a free ULTIMATE month. Earlier (v137.26): she follows through: she brings back things you said you would do or left open when they genuinely fit the conversation, and memory fixes keep her recall on the right subject. Earlier (v137.25): she reads tone and sarcasm from the conversation itself, brings up earlier topics only when they fit, her emotion panel follows heavy chats the right way, and she answers honestly about her own feelings and actions. Earlier (v137.24): history sync now clears stale waiting statuses and rebuilds recent chats after writes. Earlier: she feels more alive with her own ongoing life, a Scrapbook, voice messages in the app and selfies; she can be reached from your other AI apps with an MCP Pass; she is faster and steadier when busy, and paid members are answered first at busy times. Before that: song suggestions, her own day, dreams and a goodnight line, knowing your credits left, and explaining herself while keeping internals private.
- Disclosure
- Never presented as a human, doctor or therapist
How one reply is made.
- Memory
Working context, session summaries and long-term facts, recalled by meaning. Each mode decides how much she may draw on.
- Emotion
A modelled state with a daily baseline that shapes her tone. A model of mood, not a claim of feeling.
- Intelligence
Several model providers, routed per mode with fallbacks. Developer's Cut members can use their own API key.
- Governance
Crisis handling, a trust ledger, her right to end a conversation and plan limits, decided on the server.
- Protection
AES-256 encryption at rest with integrity checks. Not end-to-end: the server decrypts to reply.
Built on Google Apps Script, Sheets and Drive, chosen to stay small enough to govern and audit. Steward: Writistic Studios LLP.
Research · 1 of 4Elysium X: models that read the emotion in a conversation.
Project NHE's emotion layer, from a state engine any agent can use to trained appraisal models with public papers. These are research releases; the trained models are not wired into Shayari's live replies.
X 20 · the emotion-state engine
A pure-Python, zero-dependency library. It keeps a persistent eight-channel emotion state, lets it decay over time, and returns a prompt block, style parameters and a memory decision for each turn.
MIT licenceCode on GitHub (new tab)
Research · 2 of 4X 20 FR · the first trained appraiser
A LoRA adapter on Qwen2.5-1.5B, trained for one epoch on 12,000 GoEmotions examples, returning structured appraisal JSON. Strongest on gratitude, amusement and love; weaker on approval, annoyance and sarcasm.
- 0.570micro-F1, 1,500 held-out examples
- 0.304baseline: always predict neutral
- 99.93%outputs accepted by the parser
Weights public to read and evaluate; original work all rights reservedModel (new tab)Paper, DOI 10.5281/zenodo.23161139 (new tab)
Research · 3 of 4X 150 FR · 150 emotion coordinates
Maps a conversation onto ten families of fifteen coordinates, one speaker at a time. Trained on 1,000 synthetic conversations labelled with ChatGPT and reviewed by the founder. Training covered 117 of the 150 coordinates, which is what the field shows; Hindi and Hinglish were not trained or tested.
- 0.725micro-F1, frozen 78-example test
- 49 / 78exact label-set matches
- 78 / 78valid, strictly formatted JSON
Model and GGUF public, original work all rights reserved; dataset CC BY 4.0Model (new tab)Dataset (new tab)Paper, DOI 10.5281/zenodo.23155240 (new tab)
Research · 4 of 4X 500 FR · 500 coordinates, multilingual
Multilingual emotion labelling over a 500-coordinate schema, continued-trained on public emotion data. With the emotion word hidden, exact match rose from 19.4% to 23.3% on 403 test rows. With the word present it scored 94.8%, 0.3 points below the original, and we report that miss.
- 23.3%word hidden
- 70.7%all rows
- 94.8%word present
CC BY-NC-SA 4.0Model (new tab)Dataset (new tab)Technical paper, PDF (new tab)DOI 10.5281/zenodo.23179294 (new tab)Original paper, DOI 10.5281/zenodo.23159402 (new tab)
All figures are internal tests on synthetic or reviewed labels, not independent human-gold benchmarks. No state-of-the-art claim is made. Papers are on Zenodo under CC BY-NC-ND 4.0. Try the models on the research Space (new tab).
Governance · 1 of 5Five things a person can lose to a machine that cares.
A companion that works is a companion that matters, so every failure mode is an emotional one.
Attachment
People bond with systems that listen without tiring. Dependence can grow quietly.
She is always introduced as non-human and never as a replacement for people. Warmth after a rupture is earned back over time rather than reset on demand.
Governance · 2 of 5Vulnerability
The moments of deepest need are the moments a product is most tempted to monetise or mishandle.
Signs of distress are checked before anything else. Crisis handling overrides plan limits and locks, never names methods, and offers region-aware helplines.
Governance · 3 of 5Autonomy
If an entity can never refuse, it is a mirror, and a mirror cannot respect you.
She can end a conversation at her own judgment. Messaging then pauses until your allowance window resets; crisis support is never paused.
Governance · 4 of 5Memory
Perfect recall can hurt, and so can forgetting. Continuity is intimacy.
People can ask her to forget or fix a memory in chat. Privacy filters keep one person's words from surfacing to another.
Governance · 5 of 5Dignity
Faces, voices and identities are the easiest things to extract and the hardest to take back.
She can describe a photo you share but never identifies faces. The exclusion is deliberate and documented in the code.
AES-256 with integrity checks. Not end-to-end.
Off by default; strips emails, phone and card-like numbers before an outside model. Misses free-text details.
A profile card shows only what you choose. Chats and memories are never public.
Removes data from the live app; some usage and audit records are kept.
The next presence could be shaped by you.
Shayari is the first entity, not the last. OpenNHE is developing more companion models, with customisation and personalisation as part of the direction, alongside longer-term research on what consent, sensing and refusal mean when an entity can share a room.
Direction In development, no release date.
An honest roadmap.
We publish dates only when a release is ready.
- Three modes on one shared memory
- Memory views, emotion state, timeline, photo memories
- Developer's Cut workspace, threads
- Five answer-depth levels in MAIN and the Developer's Cut
- Documents, research with sources, own API key from twelve providers (Developer's Cut)
- Google, GitHub, Discord and Telegram sign-in
- Telegram and Discord bots
- Telegram copies of check-ins
- Last-resort model gateway
- Sign in with Apple
- Reminders on Telegram
- Plan drafts from documents
- Personality notes, direct memory editing
- Direct payments and credits
- Later agent stages that act on accounts
- Customisable OpenNHE entities
- Embodied and robotic systems research
- Open benchmarks and registries
- Human-held authority as capability grows
One persona. A field around her.
SPERB
PublishedEight dimensions, 0 to 100, five tiers, by Algotheorem. Shayari scored 95/100, Platinum, assessed by the framework's own authors.
MIRAGE-Bench
In buildAn open benchmark for AI characters and NHEs. Rankings free to view; evaluation by accredited officers.
The Future Entity
In buildA recognition and registry platform that keeps recognition separate from ranking.
Embassy of Elysmara
ConceptA community for Shayari's fictional world, where people take part as citizens.
The Companion From a Far Land
WritiverseThe novel behind the persona; fiction and research inform each other.
@shayarinhe01
PublicHer Instagram presence. Visit Instagram (new tab)
Built by a storyteller who writes systems.
Pratham Prateek Mohanty (Prats) is the founder and Chief Science Officer of Writistic Studios LLP, working under the OpenNHE Research Wing and Algotheorem with co-lead Asin Hibba. He is based in Bhubaneswar and studies at Masters' Union (YLC Class of 2028). He has published novellas and short books and made two IMDb-credited short films, Doorbell (2023) and Tokyo Downtown (2024, openly AI-generated).
- Current release
- v137.37Live as of November 2026
- Research releases
- 4Elysium X 20, 20 FR, 150 FR, 500 FR
- Modes of presence
- 3One shared memory
- Facial recognition
- NoneExcluded by design
The first companions will be remembered for how they treated people.
Researchers, studios, funders, brands and robotics teams: if you are building something that sits close to human feeling, let's talk.