ifp

← Garden Patch Home · Principles

• is_a::[[Principle Form]]
• has_status::[[Seed Stage]]
• in_domain::[[Deep Context Architecture]]

Zero-Tooling Floor for Knowledge Architecture

Claim

The authoring and semantic layers of the deep context architecture operate without specialized tooling. Forms are plain markdown with YAML frontmatter and predicate::[\[\[target\]\]↑](../NODES.html#:~:text=target) typed relations. Version control is git. The query layer is shell one-liners (rg 'predicate::\[\[Target\]\]' --type md). A dedicated graph-traversal skill could add value but is not required.

Scope

Applies to the authoring layer (human-facing) and semantic layer (agent-facing). The trust layer (Gordian Envelope) adds tooling requirements but is explicitly separate — the architecture works without it.

The zero-tooling floor means:

• No database
• No plugin dependency for core operations
• No schema enforcement beyond file conventions
• No build step between authoring and querying

The zero-tooling floor is also a portability guarantee. If the current tool (Obsidian, Claude Code, any future replacement) disappears tomorrow, the knowledge survives intact — readable, searchable, and structurally coherent with nothing more than a text editor and grep. No migration, no export, no data rescue. The choice of plain markdown over richer formats is a values decision: portability over power.

External Convergence

Multiple agent memory systems have independently converged on plain markdown files with YAML frontmatter as their storage format — validating the zero-tooling choice from a different direction:

• ClawVault — 8 typed memory categories as YAML memoryType: field, wiki-links for connections, folder hierarchy for organization. Benchmarks show this format outperforms purpose-built memory infrastructure (74% vs 68.5% on the LoCoMo long-context memory benchmark).
• QMD — Local markdown search engine achieving 96% token reduction by returning snippets rather than full documents. The architecture’s summary-first retrieval pattern, independently discovered.
• MIE — Cross-agent knowledge graph with typed nodes and relationships stored in markdown. 12 MCP tools for query/store/analyze across multiple simultaneous agent connections.

The convergence suggests that plain markdown with typed metadata is not merely a viable choice but the emerging default for agent-facing knowledge systems. The zero-tooling floor is also the interoperability floor — any system that can read files can participate.

Tensions

• Power vs accessibility — specialized tools (Dataview, graph databases, custom indexers) enable queries that shell one-liners cannot express. The principle doesn’t reject tools; it ensures the architecture isn’t hostage to them.
• Enforcement vs convention — without schema enforcement, structural contracts are conventions that can be violated. Path-based Claude rules provide soft enforcement during authoring, but nothing prevents manual edits from breaking contracts.

Sources

• [[Deep Context as an Architecture for Captured Reasoning]], “Authoring Layer” and “Semantic Layer” sections (lines 244-256)
• [[The Gardening Problem Returns - Document Lifecycle in the Age of AI Agents]]↑, Section 1.3 — ClawVault, QMD, and MIE as independent convergence on markdown+YAML for agent memory

Relations

• relates_to::[[Predicate Maintenance Recipes Over Tools]]↑
  • Maintenance recipes demonstrate the zero-tooling floor in practice.
• relates_to::[[Human Authority Over Augmentation Systems]]
  • Zero-tooling preserves human authority by avoiding tool lock-in.
• relates_to::[[Content Over Container]]
  • Both principles favor portable content over infrastructure dependency.

This site is open source. Improve this page.