Muninn: Living Memory System for Technical Work

Engineering teams lose critical architectural decisions, context, and dependencies between meetings and informal huddles. Muninn solves this engineering amnesia through an opt-in living memory system built on an immutable principle: "You decide what gets heard. We decide what is worth remembering." Muninn captures technical speech via a high-fidelity 24kHz AudioWorklet stream or 1-tap quick notes. Rather than dumping raw audio or rambling summaries, an edge-orchestrated intelligence pipeline breaks speech into atomic, verifiable claims classified into five durable categories: observations, hypotheses, decisions, tasks, and questions. Each claim preserves its exact conversational provenance, speaker identity, confidence rating, and turn-by-turn timestamp. These claims populate a dynamic relational knowledge graph stored in Neon Serverless PostgreSQL with pgvector. A proactive background reasoning engine continuously audits the graph for open loops, unblocked tasks, and conflicting decisions. When a blocker clears or a new decision contradicts past consensus, Muninn autonomously triggers a resurfacing alert to keep projects aligned. Muninn introduces a dual storage privacy architecture. On the web, guest users operate in an ephemeral sandbox within sessionStorage that wipes immediately upon closing the tab, with optional cloud sync. On mobile and tablet hardware, native Android storage enables local-first persistence without mandatory cloud dependence. An integrated Sensitivity Review Gate quarantines credentials, confidential financial metrics, and off-the-record statements under a strict default-discard guarantee. The entire stack is engineered for speed with zero cold starts, featuring a Next.js 15 client on Cloudflare Pages, a Hono TypeScript edge API on Cloudflare Workers, real-time speech intelligence via AssemblyAI, and cross-platform native execution via Android.