LLM Wiki — Synaptic Lattice

Nowledge Mem

Properties1
tagsentity, tool, memory, agentic, knowledge-graph, anthropic

Leitura de 4 min

Overview (confidence: high)

Nowledge Mem (mem.nowledge.co) is a cross-tool personal memory layer: you capture a conversation thread from any AI (Claude Code, Cursor, ChatGPT, Codex, browser chats), and Mem keeps the whole thread while distilling the decisions inside it into memories it can find later. It is a consumer-facing example of the agentic second brain idea — “Every AI you use. One memory that compounds.” It was surfaced in the Raindrop “Second Brain” collection.

Key Facts (confidence: high)

  • Cross-tool capture: one memory across Claude Code, Cursor, ChatGPT, Codex, and browser chats; the conversation stays whole, knowledge compounds.
  • Memory model: memories carry confidence scores and full version history; a newer conclusion supersedes an older one but the old version is kept (nothing discarded). Matches the vault’s own memory-lifecycle discipline.
  • Knowledge graph: explicit graph (example shows 2,042 nodes / 3,080 edges) with “Graph Intelligence” that walks nodes, communities, and shortest paths; also derives a live “Library” wiki (Artifacts, Entities, Crystals) from saved sources.
  • Background Intelligence: automatically links related memories, folds new evidence into old conclusions, and retires what newer evidence replaces — bounded by a “maintenance ledger” so work stays reviewed, not silent.
  • Agent access: Plugin, CLI (nmem search), auto-drafted Skills (proven against held-out checks, shipped only if they beat the current version), and MCP scoped to the task.
  • Local-first: runs offline on your machine, bring-your-own AI keys; optional Nowledge Cloud keeps background intelligence running 24/7. Free download + optional Plus subscription.
  • v0.10.0 Rust engine: Core rewritten in native Rust — native Windows and Linux support, on-device image reading, GPU acceleration, ~75% smaller download, ~33% less memory usage.
  • SKILL.md install contract: Novel onboarding pattern — a single URL (mem.nowledge.co/SKILL.md) is a machine-readable install contract. Agents self-detect their host and run the correct setup path. Supports 15+ hosts: Claude Code, Grok Build, Codex, Cursor, Gemini CLI, Copilot CLI, OpenClaw, Hermes, Droid, Alma, Bub, Pi, OpenCode, Claude Desktop, Kimi Code, Kimi Work, ZCode, MiMo Code, OMP, Proma.
  • Nowledge FS: File-system-style knowledge tree browser — exposes memories, conversations, wiki pages, working memory, activities, sources, and artifacts in one unified tree. Agents and scripts use the same layer.
  • Document Library: Drop PDF, Word, or presentation files into the Library; they are parsed and indexed alongside memories for unified search.
  • Multi-device sync: Run one Mem on an always-on machine; connect desktop, web, mobile, and tools to that same knowledge base.
  • Data portability: Export a portable text-based snapshot for backup or migration; import later without vendor lock-in.

Relationships (confidence: high)

  • agent-memory-systems — (type: belongs-to) sits in the memory-first family: persistent context across sessions, associative retrieval, continuous learning.
  • knowledge-graph — (type: uses) explicit node/edge graph with traversal and community detection.
  • memory-lifecycle — (type: compatible-with) confidence scoring + supersession with version history mirrors this vault’s lifecycle model.
  • model-context-protocol — (type: uses) exposes memory to agents via MCP, scoped to the task.
  • llm-wiki — (type: compatible-with) Mem is memory-first (not artifact-first); it complements rather than replaces the wiki pattern.

Architecture (confidence: high)

Nowledge Mem’s engine was rebuilt in Rust for v0.10.0, yielding native Windows/Linux support, on-device GPU-accelerated processing, and a dramatically smaller footprint. The system exposes four layers:

  • Knowledge Graph — explicit node/edge store with traversal, community detection, and shortest-path queries (example: 2,042 nodes / 3,080 edges).
  • Nowledge FS — tree-structured file-system metaphor over all stored data.
  • Background Intelligence — automatic link discovery, contradiction detection, evidence folding, and morning briefing, bounded by a maintenance ledger.
  • MCP Server — local HTTP endpoint that agents connect to via http://127.0.0.1:14242/mcp/ by default.

Implications

The Rust rebuild signals maturity beyond an MVP — performance and cross-platform support are now production priorities. The layered architecture (graph + FS + background + MCP) mirrors the vault’s own multi-layer design (raw → wiki → graph), confirming that successful memory systems decompose into distinct architectural layers with clear boundaries.

Implications (confidence: high)

Nowledge Mem is a real, shipped preview of the “memory that compounds across every AI you use” framing that motivates this vault. Its supersession + confidence-scored memory with full version history validates the memory-lifecycle discipline adopted here, and its cross-tool capture shows the consumer direction of the agentic second brain. The SKILL.md install contract pattern — a single URL that agents read to self-install — is a novel approach to cross-tool onboarding that this vault’s access-layer design could learn from.

Open Questions

  • How does Nowledge FS handle conflicts when the same memory is edited from multiple devices?
  • The SKILL.md install contract supports 15+ hosts — but the maintenance burden of keeping per-host connector behavior docs in sync is unknown.

Sources

^[raw/articles/raindrop-second-brain-collection-2026.md] ^[wiki/concepts/resources.md]

Visão de gráfico

Backlinks