Ecosystem

Personal assistant

Lun'Anima

listen → think → speak.

Status
In development
Version
4.4.0October 2026
Since
2024
Stack
LLM · TypeScript · Vue · Docker · SQLite
Repository

Anima · the soul

The breath at the heart of the ecosystem.

Without her, the satellites would only be tools.

A useful assistant does more than answer. It knows the context, remembers what was said and speaks up when it's worth it. Natsume follows mail, calendar and tasks, gives a brief in the morning and in the evening, and stays around during gaming sessions.

What makes the difference is continuity. Her memory isn't a message log but a structured base (facts, opinions, world model, one profile per speaker), and her mood carries over from one session to the next. She never starts from scratch.

moods
7
affinity tiers
5
voice engines
4
programs, NAS and PC
2

Architecture

How it works

Two programs. The Core runs around the clock on a NAS (state, memory, LLM, admin panel) and the Desktop Agent runs on the PC (microphone, screen, voice, avatar). A request that belongs to a satellite is delegated to it: Natsume passes it on, reports back and speaks the answer. Nothing is exposed to the internet.

Core · NAS

7 moods with natural decay, affinity over 5 tiers, absence detection

Core · NAS

Compressed short-term memory, long-term memory behind a single write gate, delegation to satellites

Core · NAS

Admin panel: any AI can be switched off, paused or triggered by hand

Desktop Agent · PC

Speech recognition (Groq Whisper) with voice activity detection and a wake word, screen vision, game hooks (FF14, Code Vein, osu!)

Desktop Agent · PC

Speech synthesis (Edge, Kokoro, Piper, ElevenLabs), RVC voice, lip-sync in VTube Studio

Where it stands

Checked in October 2026

Done

  • Core (NAS) and Desktop Agent (PC) split
  • 7 moods, affinity over 5 tiers
  • Long-term memory behind a single write gate (auto, review, off)
  • Daily consolidation and contradiction checks
  • Delegation to Lun'Acedia: mail, calendar, tasks
  • Morning and evening briefs
  • Barge-in: interrupt her mid-sentence
  • Multi-provider voice, RVC, VTube Studio
  • Admin panel with a kill switch and pause for every AI
  • Mobile access through Lun'Avaritia: pairing, topics, approval
  • Steering from the panel: paired devices, LLM spend measured without blocking
  • An inbox in the panel that acts on the real source

In progress

  • Real-world testing on the NAS
  • Response latency measurement
  • Memory records per person and per topic

Next

  • Personality overhaul