LLM wiki engine for R. Your knowledge base grows while you work.
An open source R implementation of the Obsidian markdown vault concept combined with Karpathy's LLM wiki pattern. The LLM maintains the wiki. You edit, curate sources, and ask questions.
pensar ("to think") has one dependency: yaml.
remotes::install_github("cornball-ai/pensar")Most people's experience with LLMs and documents is stateless. You ask a question, the LLM retrieves some chunks, generates an answer, and forgets everything. Nothing compounds.
pensar takes a different approach. An LLM reads your sources (articles, chat logs, project briefings, whatever you point it at), writes wiki pages that synthesize the key ideas, links everything together with [[wikilinks]], and maintains an index. The knowledge base gets richer with every source you add and every question you ask.
The vault is the synthesis layer, not a data lake. Sources that already live somewhere (corteza sessions, saber briefs) get referenced, not copied. Content without a home (a link someone sent you, a pasted article, a quick note) can be ingested directly.
library(pensar)
# Create a vault (pick a path you actually want to look at)
init_vault("~/wiki")
use_vault("~/wiki") # persist for the session; add to ~/.Rprofile
# Ingest a source
ingest("Article content here...",
type = "articles",
source = "https://example.com/interesting-post",
title = "Interesting Post")
# Check what's in the vault
status()
# Find what links to a page
backlinks("Interesting Post")Vault path resolution: vault argument → options("pensar.vault") (set by use_vault()) → PENSAR_VAULT env var → tools::R_user_dir("pensar", "data"). The R_user_dir() fallback is CRAN-safe but lands in ~/.local/share/R/pensar/, which is unpleasant to live in. Pick a real path with init_vault("~/wiki").
{vault}/
raw/
articles/ clipped articles, pasted text, links worth preserving
chats/ conversation logs worth keeping
briefings/ project briefings (one per project, historical record)
matrix/ messages from Matrix rooms
wiki/ LLM-maintained pages (summaries, concepts, analyses)
index.md auto-generated catalog of everything
log.md append-only record of operations
schema.md conventions for content in the vault
CLAUDE.md instructions for Claude Code when started here
AGENTS.md same content for Codex and other agents
{name}.Rproj RStudio project file
raw/ is for content you want to preserve in the vault. ingest() stores it there. Sources that already live somewhere and don't need preservation can be referenced directly by wiki pages in their frontmatter, no ingest needed. index.md and log.md are maintained by pensar functions.
init_vault() seeds CLAUDE.md and AGENTS.md by default so any agent you start in the vault (Claude Code, Codex, etc.) knows how to operate on it — what files are immutable, how to drill down with pensar show, when to rebuild the site, and so on.
For conversational use, start your agent session in the vault directory itself. The working directory becomes the knowledge base, auto-memory stays scoped to vault work, and file edits land in the right place by default.
Pass agent_instructions = FALSE to init_vault() if you don't want these files.
Use both, for different things:
- Git for the vault source (
raw/,wiki/,index.md,log.md,schema.md, etc.). The vault is plain markdown — it diffs beautifully, history matters when a wiki page gets revised, and you can push to a private GitHub repo for backup. Afterinit_vault(), just rungit init && git add . && git commit -m "initial vault". - Syncthing (or Dropbox, etc.) for the rendered site (
vault_export()output), so you can browse on your phone without running anything. SetPENSAR_SITE_DIRto a synced folder andpensar exportwrites there by default.
Don't sync the vault source via Syncthing. Concurrent edits from multiple devices on the same .md file get messy, and you lose history. Use git for that.
Note: "raw" in pensar terminology means source documents in raw/ (vs. synthesized wiki/ pages), not "raw text". Everything in the vault is markdown — there's no separate raw-vs-rendered distinction inside the vault itself. Rendering happens via vault_export() into a separate directory.
| Function | What it does |
|---|---|
init_vault(path) |
Create the vault directory structure and seed control files |
use_vault(path) |
Remember a vault path for this session (sets options("pensar.vault")) |
ingest(content, type, source) |
Write a source to raw/, update the index and log |
update_index(vault) |
Regenerate index.md from all vault pages |
log_entry(message, operation) |
Append a structured entry to log.md |
status(vault) |
Page counts by category |
backlinks(page, vault) |
Find all pages linking to a given page |
outlinks(page, vault) |
Find pages this page cites |
show_page(page, vault) |
Content + outlinks + backlinks for drill-down |
lint(vault) |
Orphans, broken wikilinks, tag clusters needing synthesis |
vault_export(vault, out_dir) |
Render vault to static HTML (requires pandoc) |
A pensar CLI is also installed at {pkg}/bin/pensar:
pensar status page counts by category
pensar lint health check
pensar show "<page>" drill-down inspection
pensar back "<page>" backlinks only
pensar tag <tag> pages with this tag
pensar log [n] last n log entries
pensar export [out-dir] render to static HTML
Symlink {pkg}/bin/pensar to somewhere on your PATH (e.g., ~/.local/bin/pensar) to use it as a command.
Every page uses YAML frontmatter and plain [[wikilinks]]. Compatible with Obsidian but no Obsidian-specific extensions. View the vault in RStudio, Obsidian, any markdown editor.
---
title: Page Title
type: concept
source: https://example.com
date: 2026-04-13
tags:
- R
- testing
---
This connects to [[Other Page]] and builds on [[Raw Source]].| Package | Purpose |
|---|---|
| saber | AST symbol index, blast radius, package introspection |
| pensar | LLM wiki engine (this package) |
| corteza | Agent runtime and chat loop |
| llm.api | LLM provider connectivity |
| mx.api | Matrix Client-Server API |
Apache 2.0