The raven that remembers.
Muninn is a personal assistant you talk to by message, terminal, browser, or voice — and it keeps one shared memory across all of them. It doesn't grow a giant context window; it remembers. Facts are pulled from every conversation, merged, and faded over time, so it holds onto what matters and lets the noise go.
Small context, deep memory
Instead of stuffing everything into the prompt, Muninn keeps a real memory and pulls in only what's relevant to what you just said.
- Extract. After every exchange, a background pass pulls out the durable facts and stores them — quietly, off the critical path.
- Consolidate. Similar facts are merged, contradictions are corrected, and a periodic "dream" pass re-reads recent history for higher-order connections.
- Decay. Memories fade on a curve unless they're used again — so the store stays sharp instead of drowning in stale detail.
- Recall. Each message triggers a hybrid search (keyword + meaning) over the store; a small model then judges which hits are actually relevant before they're injected.
One router, many hands
The model in the conversation sees a single tool. Behind it, a sub-agent wields the real toolset — so capabilities grow without bloating context.
▲ Knows who's talking
An identity layer bridges chat handles, voice prints, and devices into one person, with per-person permissions and a careful path for unknown speakers.
▲ Runs the house
Reads and controls the smart home in plain language — lights, climate, scenes — and answers "is anything still on?" Includes an opt-in open-house mode for guests.
▲ Music & a radio DJ
Full multi-room music control, plus an opt-in DJ mode: a voice that speaks a short, language-matched intro over each new song, grounded in real artist facts.
▲ Presence & a family map
Knows who's home and who's together, and feeds a live people layer and custom places to a self-hosted map — without ghosting on a stale GPS fix.
▲ Mail, calendar & contacts
Reads and sends email, keeps a shared calendar with reminders it actually delivers, and resolves "email so-and-so" against an address book.
▲ Deep research
One ask becomes a cited brief: it plans, searches, fetches, compresses, and synthesizes — bounded and runnable in the background.
▲ Sees & speaks
Describes a camera view on request, generates images, and talks back through natural text-to-speech — all on local GPUs.
▲ Resilient & self-aware
A backup brain and load-sharing keep it alive if one model host drops; it logs its own failures and distils them into rules to avoid next time.
From a spoken word to an answer
Small satellites around the house listen all the time. A gate on the GPU server decides which sentences are meant for Muninn, and the answer comes back out of the same room it was asked in. The tablets run Bifrost.