An LLM Wiki grows smarter over time through two distinct learning modalities: continual learning (temporal accumulation and retention) and recursive learning (iterative self-application and refinement). Both are essential to the Vivary agent loop and the Wiki CLI ecosystem, but they operate at different time scales and target different dimensions of the knowledge base.

Continual learning

Continual learning is the forward accumulation of knowledge over time: each session, each document, each query adds to the corpus without catastrophic forgetting. The wiki grows as a compounding asset — pages are added, frontmatter is enriched, the RDF graph expands — and prior knowledge remains accessible and valid.

In this ecosystem, continual learning manifests as:

• The LLM Wiki pattern — a persistent, interlinked Markdown corpus that an agent maintains across sessions. Raw inputs are compiled into structured pages; the knowledge base compounds rather than being regenerated each time.
• Farzapedia — Farza's proof-of-concept: 2,500 unstructured entries became ~400 clean, linked wiki articles, then continued to grow and improve.
• Wiki CLI validation — wiki check and wiki lint ensure each new page meets structural and convention standards, so the corpus stays healthy as it accumulates.
• Agent Memory Filesystems — persistent memory stores (SMFS, MemFS, Wiki CLI) that survive session boundaries and let knowledge accrete.

The key property: knowledge persists and compounds forward. The agent does not start from zero each session.

Recursive learning

Recursive learning is iterative self-application: prior outputs, learned structures, or verified states feed back into the next cycle as inputs. The agent refines its own process by examining what it produced and improving it.

In this ecosystem, recursive learning manifests as:

• The Vivary strato loop — Ask → retrieve → act → verify → learn → gate. The "learn" step examines verification results and the gated outcome, then distills improvements into memory, skills, or the graph itself.
• Vivary ozone — graph-aware review that checks not just individual documents but the relationships between them, surfacing gaps a per-document check cannot see.
• Procedural Knowledge — self-updating workflows: SPARQL blocks that render live results, SHACL shapes that validate structure, and wiki skills that encode repeatable processes.
• Wiki Subcommand render — wiki render --check detects stale SPARQL result blocks from a prior run and flags them for regeneration, closing the recursive loop.

The key property: the system improves its own process by examining its output. The loop feeds into itself.

Comparison

How they compose

Continual and recursive learning are complementary, not competing. A healthy agent-managed wiki uses both:

1. The agent adds and refines pages over time — continual growth of the knowledge base.
2. On each turn (or heartbeat), the agent verifies its latest output, learns from verification results, and updates its skills or memory — recursive refinement of process.
3. The recursive loop feeds the continual corpus: insights from verification become new frontmatter, corrected links, or updated shapes.

The LLM Wiki pattern is the continual surface; the Vivary strato loop is the recursive engine. They meet in the middle: a wiki that grows perpetually and improves itself on every interaction.

Related

• LLM Wiki — pattern origins and compounding knowledge base design
• Vivary — agent workspace with the strato loop and ozone review
• Agent Memory Filesystems — persistent memory across sessions
• Declarative Knowledge — facts and structures that accumulate
• Procedural Knowledge — workflows and processes that self-improve
• Farzapedia — continual learning proof of concept
• Wiki Subcommand render — stale-block detection as recursive feedback