Memoriki: A Memory Layer for Claude Code

Memoriki is a personal knowledge base project that pairs two ideas: Andrej Karpathy’s LLM Wiki pattern and MemPalace, an MCP server. The repository, published at https://github.com/AyanbekDos/memoriki, summarizes the combination plainly: “Wiki gives structure. MemPalace gives memory.” The goal is to give a coding agent such as Claude Code a place to store and retrieve knowledge that persists outside a single conversation.

The Problem It Addresses

The README frames the limitation of using either approach alone. An LLM wiki by itself is described as a “library without a catalog,” because searching it amounts to running grep across files. MemPalace on its own is a “search engine without books.” Put together, the project aims to provide structured knowledge, semantic search, and an entity graph in one setup.

Three Layers of Knowledge

Memoriki organizes information into three layers, each with its own role.

The Wiki layer supplies structure through Markdown files, wiki-links, YAML frontmatter, and an index, and is meant to be browsed in a tool like Obsidian. The MemPalace Drawers layer adds semantic search over content, queried through a mempalace_search tool. The MemPalace KG layer holds an entity relationship graph with timestamps, queried through mempalace_kg_query.

The repository keeps source material in a raw/ folder for articles, notes, and transcripts. A separate wiki/ folder is owned entirely by the language model and contains index.md, log.md, and subfolders for entities, concepts, sources, and synthesis. Configuration lives in mempalace.yaml, while the schema and rules are recorded in a CLAUDE.md file.

Setup

The quick start in the README walks through a short sequence. A developer clones the repository into a knowledge-base folder, installs MemPalace with pip install mempalace, and runs mempalace init . to initialize the directory.

Connecting it to Claude Code uses the MCP command:

claude mcp add mempalace -- python -m mempalace.mcp_server

After that, a source file is dropped into the raw/ folder, Claude is launched, and the agent is asked to ingest the file into the wiki.

How It Is Used

The project describes three core operations. Ingest means dropping a file into raw/ and having the model integrate it into the wiki. Query means asking a question, with the model finding the relevant pages and synthesizing an answer. Lint is a health check that looks for contradictions, orphaned pages, and knowledge gaps.

Compatibility and Licensing

Because Memoriki is built on MCP, the README states it works with any MCP-compatible agent. Claude Code uses the included CLAUDE.md file as written. For OpenAI Codex, that file is renamed to AGENTS.md. Cursor and the Gemini CLI are also listed as compatible. The project is released under the MIT license.

Memoriki is an early effort, so developers evaluating it should treat the repository itself as the authoritative reference and check it for updates before relying on the workflow for important projects.