memm.dev

What It Is

MEMM is positioned as a “second brain for AI.” It is not a chatbot or a large language model, but a local, portable, MCP-native AI memory and context engineering tool. It stores knowledge as Markdown files with YAML frontmatter, and serves relevant context to ChatGPT, Claude, Cursor, Codex, and local LLMs through a local MCP Server. The goal is to reduce repeated explanations of project background and wasted tokens.

Core Capabilities

MEMM’s core strengths are filesystem-native memory, typed ontology, and scored retrieval. Memories can be labeled as Entity, Concept, Source, Synthesis, and other types, with different levels of detail exposed through L0/L1/L2 layers. On the retrieval side, it scores content using six types of signals: BM25, semantic relevance, graph relationships, recency, importance, and access frequency, then injects the most relevant content based on the token budget. It also provides governance features such as health checks, expired memories, conflict detection, and merge suggestions. One caveat: the documentation indicates that current semantic retrieval is closer to keyword expansion and phrase overlap, while true local embeddings remain part of the roadmap.

Pricing and Privacy

The page clearly states that MEMM is free and open source. Its source code is available on GitHub, licensed under FSL and automatically converting to Apache 2.0; no commercial edition or subscription pricing is disclosed. Privacy is one of its strengths: there is no cloud sync, telemetry, or analytics. The MCP Server binds only to 127.0.0.1, and knowledge is stored as local Markdown files that are readable, editable, and manageable with Git.

Pros and Cons

The advantages are transparency, auditability, no vector database infrastructure burden, and the ability to reuse the same memory across tools. Compared with manually maintaining CLAUDE.md, it is more structured and maintainable. The drawbacks are that the product is still early: the page marks it as v0.1.0, and the public Mac version is still under development. Users need to understand Markdown metadata, types, and governance workflows, so there is a learning curve. The official page cites metrics such as +43% accuracy and 87% token savings, but the main text does not provide sufficiently complete evaluation details.

Who It’s For and Access from China

MEMM is best suited for engineers, indie developers, and small teams, especially for capturing long-term project knowledge such as architecture decisions, coding conventions, error-handling rules, source materials, and postmortems. The page does not state how well it can be accessed from China. As a local application, it should theoretically run without relying on cloud services, but downloading from GitHub and connecting to external tools such as ChatGPT or Claude may be affected by network and account availability. Alternatives include Obsidian/Logseq, manually written AGENTS.md/CLAUDE.md files, or vector RAG solutions such as Chroma and pgvector.