It started with a 1987 idea.
It became something of its own.

The 1987 inspiration

The concept film was five minutes long. A touchscreen tablet sat on a professor's desk. An agent on the screen watched his calendar, retrieved papers from a remote library, placed a call to a colleague, navigated an interruption from his mother, and helped him finish a lecture. It wasn't a chatbot. It was a person you trusted with your day.

The first scene goes something like this: the professor asks for last semester's lecture notes — specifically the ones with the deforestation animation. The agent finds them. Three quiet specifications hide in that one request:

• Persistent identity. The agent already knows who "I" am — it doesn't need a login each turn.
• Long-tail memory. "Last semester" is a real time anchor. The agent recalls a specific lecture out of dozens.
• Multimodal grounding. The deforestation animation is recognised because the agent has both seen it and indexed it semantically — not just by filename.

The 39-year gap

From 1987 to 2024, none of the building blocks were quite ready. Speech recognition was brittle until 2014. Useful on-device language models didn't exist until 2022. The hardware to run them on a phone or a laptop without burning a hole through your battery is only a few years old. And the cultural pattern that does exist — the chatbot — is the polar opposite of what 1987 imagined: stateless, server-side, no memory, no voice, no awareness of the person.

The vision didn't die because it was wrong. It waited because the pieces weren't shippable. They are now.

What's different now

Three platform shifts in the last four years made the 1987 vision achievable for a small studio, not just a moonshot for a hyperscaler:

M-series chips mean real local AI

M-series Mac chips give every modern Mac a unified-memory GPU and a Neural Engine that can run small language models, ASR, and embeddings at interactive speed. The same kind of silicon sits in every recent iPhone — and every Vision Pro, Watch, and TV. The compute is already on the user's side of the wire.

Useful small LLMs exist

On-device Foundation Models, Llama 3, Mistral, Phi, and Qwen variants run locally and are good enough for the bread-and-butter agent tasks: classification, extraction, short-form chat, reranking. We don't have to ship every token to a hyperscaler to get something useful.

The cultural appetite is here

Three years of cloud chatbots taught a lot of people that sending every keystroke to a model farm in Virginia is not the only path. Public posture on data ownership has shifted; "build it on my device, respect what I tell you to forget" is now a sellable promise, not a niche request.

Four pillars, in detail

The home page lists them in one sentence each. Here's what each one actually means in the product.

① Memory you can see

Every fact bimi remembers is a typed record with provenance, confidence, and timestamps. Identity facts ("Steven works at TBC Corp"), preferences ("doesn't drink coffee after 2pm"), commitments ("emailed Maya on Tuesday to set up a 1:1"), and episodes ("the dinner on the 14th of May") are categorised separately so the surface in the Memory Inspector reads like a sky: stars by category, not a JSON dump.

Each star has an edit affordance, a "forget" affordance, and an audit trail. There is no privileged path through which bimi can know something about you that isn't visible in the constellation.

② On-device first

bimi's default routing is local. On-device Foundation Models handle anything that fits in their context window, including classification, extraction, and short-form chat. MLX and llama.cpp pick up the slack for larger models on M-series chips. The cloud is opt-in, listed by provider, and shown in the routing decision.

When something does spill to the cloud — long-form reasoning, embedding a large corpus, working with a hosted multimodal model — the routing screen names the provider, the model, and the reason. No silent escalations.

③ Privacy is the posture, not a setting

Personal-sensitive facts are encrypted with your device key and stay on the device by default. Nothing about you is persisted in a place you can't inspect, edit, or delete from the constellation. There are no third-party analytics SDKs in the binary; crash diagnostics attach only to bug reports you choose to send.

The vision in 1987 was an assistant that knew you intimately without that knowledge ever leaving your control. That's the target.

④ Proactive, not pushy

bimi watches signals — calendar events, location pings, app activity, music — and promotes the salient ones into memory silently. Reflection runs between turns, distilling patterns. When it speaks, it has a reason. When it doesn't, it stays quiet. The opposite of a notification you wish you could turn off.

Why most chatbots fail this vision

Mainstream chatbots — ChatGPT, Claude.ai, Gemini — are extraordinary general-purpose tools. They are also structurally unable to be the 1987 vision. Four reasons:

1. No persistent identity. They greet you fresh every conversation. "Memory" features bolt on a chunk of text the system reads at the start of each chat. There's no constellation, no audit, no decay.
2. No on-device default. Every token, every voice clip, every screenshot leaves your device by design. There is no privacy mode where your machine does the work and the cloud sits on the bench.
3. No native surface. The product is a webview. It can't watch a calendar event happen, notice what's playing, read the config file you just dropped, or speak when something matters.
4. No restraint. The interaction pattern is "I ask, it answers, please rate this answer." The 1987 vision's pattern was "it observes, it speaks once when it should." Different shape entirely.

We're not saying those products are bad. We use them. We're saying they can't be the kind of agent the 1987 video imagined. Someone has to build the other thing.

How bimi maps to the vision

Where a bimi decision does echo a scene from the 1987 video, the mapping is worth showing. Plenty of the product has no line in that film at all — the table below is only the part that rhymes.

For the long version, head to the roadmap — every feature is broken out with status, the open architectural questions, and what depends on what.