The bibliography
behind the wiki
The gist that named the pattern, and the memory and context papers around it. This page says what each one argues and which part Paniolo ships. Commands stay on the wiki guide in the nav above.
Motivation, not a score
None of these sources measured paniolo wiki. Karpathy's note is an idea file. The papers measure memory systems, written context, or agent files in their own setups. Where Paniolo takes a shape from them, the paragraph says so. Where a paper's mechanism is a neighbor we have not shipped — automatic note rewriting, query-driven restructuring — the paragraph says that too. The 2026 measurements of compiled wikis, trained walkers, and question-shaped stores are on knowledge base research.
Compile knowledge
once, then keep it
LLM Wiki — Andrej Karpathy
The gist splits a knowledge base into three layers. Raw sources stay immutable: the model reads them and does not edit them. The wiki is interlinked markdown the model owns — entity pages, concepts, syntheses, an index, and a log. The schema is a short conventions file (AGENTS.md or CLAUDE.md) the model writes against. The operations are ingest a source, query the wiki, and lint for drift. Karpathy's line is that the editor is the IDE, the model is the programmer, and the wiki is the codebase. It is written as something you hand to your own agent, not as a study with a result.
Paniolo's wiki command is that instantiation. paniolo wiki init stamps the conventions, index, and log. paniolo wiki lints page shape, frontmatter, wikilinks, and near-duplicates. Page operations — new, rename, move, status, archive, delete — are the bookkeeping, and they are explicit: the destructive ones plan until you pass --apply. Retrieval is paniolo qmd, so a question searches the compiled pages instead of re-reading the raw sources from scratch. The retrieval bibliography is on qmd research.
Karpathy's LLM Wiki as Agent Memory — Angie Jones
Jones maps the same pattern onto the usual agent-memory taxonomy and argues it covers the long-term types without a special memory server. Semantic memory is the concept and synthesis pages. Entity memory is one page per named thing. Episodic memory is the log of what was ingested, decided, or superseded. Summary memory is each source compressed onto its page. Procedural memory is the schema the agent writes against. Conversational memory and working memory stay inside the session; they are not what the wiki is for.
That split is how a Paniolo wiki is filed. log.md is the episode record, written by wiki new, status, and archive. Kind prefixes (plan-, decision-, paper-, and the rest) are the filing scheme the schema demands. The always-loaded conventions stay in the schema file. The pages themselves are searched, not pasted into every prompt. Two memory types Jones leaves in-session — the chat and the scratchpad — Paniolo also leaves in-session.
What sits beside
the pattern
A-MEM — Agentic Memory
A-MEM treats memory as a Zettelkasten: when a new note arrives, the system links it to older notes and can rewrite those older notes so the store stays a living network rather than an append-only log. That is the automatic-evolution fork of the same design space Karpathy describes by hand. The paper is about that memory system in its own tasks. It is not a measurement of a linted markdown wiki.
Paniolo keeps the linked pages and refuses the unattended rewrite. A new page is wiki new. A link change is rename or move, which rewrite references because you asked. Nothing in paniolo qmd edits a page because a query arrived. If the corpus should change shape, an agent does it as a page operation you can review, then the linter checks what it touched.
Codified Context
The production study behind this paper followed one always-loaded constitution — about 660 lines, checked across 283 sessions — on a codebase of 108,256 lines. The context infrastructure around it was 24.2% of the repo. Those figures are the paper's, from that deployment. They are also why paniolo scan treats always-loaded guidance as a budget, with a default ceiling tighter than 660 lines for smaller repos. The full trace is on the scan research page.
The wiki is how that budget stays small. The schema file is the constitution: short, always loaded, and linted as a conventions file. Decisions, plans, and source notes live as wiki pages and come back through search when a task needs them. Stuffing the corpus into AGENTS.md would spend the budget the paper shows is already large. paniolo wiki and paniolo qmd are the split: a thin always-loaded contract, and a compiled store beside it.
DeepRefine
DeepRefine studies query-driven refinement: the wiki is not only written when a source arrives, it is restructured because of the questions people ask. Passages that queries keep missing, or questions the current outline cannot answer, become a reason to rewrite the corpus. That is a real neighboring idea. It describes a refinement loop. It is not a static linter.
Paniolo does not restructure a wiki from the query log. A search that misses is a retrieval result, reported by qmd, not a license to rewrite pages. The maintained artifact still changes when someone ingests a source or runs a page operation. Query-driven rewrite is the part of this paper we have not shipped. What we did ship is the other half of Karpathy's loop: lint the pages you have, and search them, so a miss is visible before anyone edits.
DeepRefine — arXiv 2026, arxiv.org/abs/2605.10488Meta-Harness
Meta-Harness puts agent memory in the filesystem and uses it to search a harness, arguing for the full history of what was tried over a summary that forgets the path. The useful claim for a wiki is modest: durable files beat a summary that lives only in a chat, and search over those files is how an agent consults them. The paper's subject is that harness-search setup.
A Paniolo wiki is that filesystem. Pages are ordinary markdown. log.md keeps the ingest and archive history instead of collapsing it into the latest summary. wiki archive moves a finished page to raw/wiki-archive/ and repoints links, so the history remains a file qmd can still be pointed at, while the live corpus stays the pages an agent should treat as current. The harness scan is a different command; this paper is why the wiki is files and a log, not a hidden vector store you cannot open.
Externalization in LLM Agents
This review lines up four places agents put capability outside the weights: memory, skills, protocols, and the harness around the model. The wiki is a memory externalization. A skill is a procedure the agent can be handed. A protocol is how tools and people agree to act. The harness is the operating environment — guidance, checks, and the loop that runs them. Mixing those four into one file is how context files grow until nothing in them is reliable.
Paniolo keeps the four apart. The wiki is the memory store, linted and searched. Skills are a separate catalog, installed with npx skills, not pasted into wiki pages. paniolo scan scores the harness: guidance, guardrails, and wiring. The review is why a wiki page that starts containing install steps, tool policy, and a decision log is a lint problem, not a feature. Each of those belongs in the layer that owns it.
Evaluating AGENTS.md
This paper asks what an AGENTS.md-style context file actually changes in agent performance, instead of assuming that more written instructions are better. The result that matters here is the question, not a number we can carry over: a context file is an intervention with a size, a placement, and an effect, and those can be studied. Paniolo has not reproduced the paper's evaluation on customer repos.
In the wiki, that file is the schema, not the knowledge base. wiki init stamps a short conventions file the linter and the page operations assume. The decisions and the source notes go in pages. paniolo scan then measures how much guidance is always loaded, which is the product consequence of treating AGENTS.md as something whose size should be checked rather than grown without a limit.