Skip to content
Erebus edited this page Sep 30, 2026 · 3 revisions

Welcome to the Muninn Wiki

Muninn is an opt-in living memory system for engineering teams. It captures technical discussions (via real-time 24kHz AudioWorklet stream or 1-tap quick notes), isolates verifiable ground-truth claims with turn timestamps, weaves a relational knowledge graph, and proactively resurfaces unfinished work, unblocked tasks, and conflicting decisions.


Quick Navigation

flowchart TD
    WikiHome["Muninn Wiki Hub"] --> Arch["System Architecture"]
    WikiHome --> Voice["Dual-Mode Voice Agent"]
    WikiHome --> Graph["Knowledge Graph & Resurfacing"]
    WikiHome --> API["OpenAPI & Route Reference"]
    WikiHome --> Privacy["Dual Storage & Privacy"]
    WikiHome --> Deploy["Getting Started & Deployment"]

    Arch --> ArchDesc["24kHz Pipeline, Edge Workers, Neon pgvector"]
    Voice --> VoiceDesc["Co-Pilot vs Scribe, VAD, WebSocket contracts"]
    Graph --> GraphDesc["5 Atomic Claims, CTEs, Cascade Deletions"]
    API --> APIDesc["Sessions, Claims, Review, Resurfacing, Chat"]
    Privacy --> PrivacyDesc["Ephemeral Web vs Android Flash, Review Gate"]
    Deploy --> DeployDesc["Cloudflare Workers, Pages, Android APK"]
Loading

Core Philosophy

  1. Capture is opt-in, per session: No ambient or always-on listening. The user initiates a Muninn session explicitly, like starting a voice memo or architecture huddle.
  2. Triage happens inside the session: Not everything said is equally durable, and some content is sensitive. Triage is a post-hoc filter on consented data:

    "You decide what gets heard. We decide what is worth remembering."

  3. Core Cognitive Loop:
    TALK -> UNDERSTAND -> CONNECT -> REMEMBER -> NOTICE SOMETHING IMPORTANT -> TELL ME
    
  4. Verifiable Provenance: Every stored claim retains its origin (source conversation, speaker, turn timestamp, and confidence status).
  5. Zero Cold-Start Edge Execution: The primary backend runs on Cloudflare Workers with Hono and Neon Serverless PostgreSQL (@neondatabase/serverless), guaranteeing sub-millisecond response times without container spin-down delays.

Live Deployments & Key Links