Author: Caspar Addyman Published: 2026 (personal blog / newsletter) Source: Readwise highlights
Summary
Addyman reflects on what running an Obsidian vault to ~18,000 files has taught him about where personal knowledge systems are going. The thesis: a specific combination of components — local AI for private semantic search, visual thinking surfaces, structured database functionality, and external AI integration — has just crossed a threshold where it enables qualitatively new patterns of personal knowledge work. He frames it as a window that did not exist six months prior. The implicit argument: PKM scale (tens of thousands of notes) is no longer a problem to solve, it is a feature to exploit, provided the right stack is in place.
Key Takeaways
- Scale is now an asset, not a liability. "For the first time, we can build personal knowledge systems that scale to tens of thousands of notes without breaking, find patterns across years of accumulated knowledge, integrate with AI while keeping data private, support both analytical and creative thinking, and adapt and evolve with new tools and workflows."
- Local AI is the privacy-keeping core. "Smart Connections keeping everything private" — semantic similarity over the vault without the vault leaving the device. This is the technical fact that makes large-scale personal AI compatible with sensitive notes.
- The combination is the point, not any one tool. Local AI + visual thinking (Canvas, Excalidraw) + database functionality (Bases) + external AI (MCP) "creates possibilities that didn't exist even six months ago." None of these alone is the breakthrough; the integration is.
- Smart Connections enables forgotten-knowledge recovery. "Smart Connections to find patterns across 5 years of notes I've forgotten about." Local semantic search closes the gap between capture and recall — the Knowledge Decay problem partially solved.
- The pattern is reproducible. Addyman's stack is not bespoke — it is composed of widely available Obsidian plugins and standards. The future of digital knowledge is not closed-platform "AI-native PKM apps"; it is composable open-source tooling that anyone can assemble.
Concepts Mentioned
- Vault scale (~18,000 files) as a threshold, not a limit
- Local AI / on-device semantic search (Smart Connections)
- Visual thinking surfaces (Canvas, Excalidraw)
- Database functionality (Obsidian Bases)
- External AI integration (MCP)
- Pattern discovery across years of notes
- Privacy-preserving AI integration
- Composable PKM stacks
- Adaptive workflows
- The "six-months-ago" frame for capability windows
Entities Mentioned
- Caspar Addyman (author)
- Obsidian (the substrate)
- Smart Connections (Obsidian plugin — local AI semantic similarity)
- Canvas, Excalidraw (visual thinking plugins)
- Bases (Obsidian database functionality)
- MCP (Model Context Protocol — external AI integration standard)
Relevance to PKM
This article is one of the strongest pieces of empirical evidence for The 2025-2026 Moment: a specific window where preconditions converged. It also operationalizes Local-First and Data Sovereignty beyond philosophical principle — Addyman is running real workflows that depend on local AI to keep tens of thousands of notes private while still benefiting from semantic search.
It directly informs PKM Stack — the specific combination Addyman describes (Obsidian + Smart Connections + Canvas + Bases + MCP) is a reference architecture, not an arbitrary toolset. It also answers the scale question implicit in Personal Knowledge Management: yes, vaults can scale to five-figure note counts without collapsing — but only if the retrieval and integration layers are in place.
The privacy framing matters: Addyman's stack is not local-AI-or-cloud-AI; it is local-AI-as-default-for-private-content with cloud-AI-as-optional-extension. That is a richer privacy architecture than blanket cloud-avoidance.
Open Questions
- How does the workflow change at 50,000 or 100,000 files?
- What is the practical accuracy ceiling of local embedding models against cloud frontier models, and is the gap closing?
- When external AI integration via MCP is used, what is the privacy boundary — what leaves the vault and what stays?
- Does the Addyman stack require a specific cognitive style (visual + analytical) or generalize?
References
- Original article: Readwise capture (Addyman, 2026)
- Smart Connections plugin: smartconnections.app
- Obsidian Bases, Canvas, MCP — see PKM Stack