LLMs need wikis, not just chat history

Chat history is a convenient memory format for a conversation. It is a weak memory format for a project.

Projects have names. They have folders. They have source documents, notes, decisions, queries, tables, and reports. They have facts that need citations and claims that need to be updated when the source changes. A transcript can mention those things, but it is rarely the right place for them to live.

LLMs need access to the workspace, not only the residue of prior conversations.

Chat is chronological, work is addressable

The basic unit of chat history is “what happened next.” The basic unit of a workspace is “where does this live?”

That difference matters. An agent trying to answer a question needs stable addresses:

• research/q3-report.pdf
• notes/vendor-evaluation.md
• analysis/revenue.sql
• reports/board-brief.md

Paths let a human and an agent point at the same thing. Wiki links add another layer: a note can say [[vendor-evaluation]], and the workspace can resolve that into a resource. The reference is not trapped inside one chat turn.

Wikis make context explicit

A wiki-shaped workspace gives the model structure before retrieval starts. Links say which documents are related. Embeds say which views belong inside a note. Headings divide a document into addressable sections. Citations tie claims back to evidence.

That structure does not need to be fancy. It can be plain Markdown:

The renewal risk is highest for [[accounts/acme]].

Evidence: [Q3 procurement note](raw/q3-procurement.pdf?cite=node_123)

The important part is that the structure is visible. A human can inspect it. A search index can use it. An agent can follow it. The same object can survive outside the product that rendered it.

The source of truth should not be a vector store

Vector search is useful. It is not a workspace.

If the only durable representation of knowledge is an embedding index, users cannot browse it, edit it, diff it, or hand it to a different tool. They can ask questions against it, but they cannot really own it.

Mempipe treats search as a derived layer over files. The files remain the source of truth. The index helps find evidence. The answer cites the evidence. The citation opens the source. That loop is much easier to trust than an answer floating above an opaque memory system.

Agents should write artifacts, not only replies

A useful agent should leave behind work the human can reuse. Sometimes that is a Markdown brief. Sometimes it is a saved SQL file. Sometimes it is a cited table or a note linked into the rest of the workspace.

This is where wiki-shaped files matter. The output is not just “the assistant said X.” It becomes a named artifact with links and citations:

# Vendor risk summary

Acme has a renewal risk tied to procurement timing. See [[accounts/acme]] and the cited procurement note below.

Now the next human, tool, or agent can pick it up without replaying the whole conversation.

Memory should compound outside the session

The right long-term memory for LLM work is not a longer transcript. It is a workspace where important context gets promoted into durable files.

Chat stays useful for interaction. Files stay useful for memory. Wiki links connect the files. Citations keep the claims accountable. Search and retrieval make the whole thing fast enough to use.

That is the shape Mempipe is building toward: not a chatbot with a bigger memory, but a file-native workspace that agents can actually use.

The concrete form of this is an LLM wiki: raw sources in one folder, agent-maintained Markdown pages in another, and a schema file that says how to keep them in step. The how-to builds one from a PDF, with citations that open the page.