Memex adds local long-term memory for Claude users

Memex is a new tool that gives Claude persistent memory across chats using local RAG and offline embeddings. The project stores notes on the user’s machine, works with Obsidian vaults, and does not require cloud services or API keys.

Memex adds local long-term memory for Claude users

Memex is a new local memory layer for Claude that stores notes on the user’s machine and surfaces them later through semantic search. According to the project description on Show HN, it is designed to act as a “second brain” for Claude, giving the assistant persistent memory across conversations without sending data to a cloud service.

The tool is installed with npm install -g @evan-moon/memex, and the project says it keeps all data on local storage at ~/.memex/memex.db. Notes are saved as Markdown files, while embeddings are stored in a local SQLite database. The developer says nothing leaves the user’s machine.

⚡ New to this?

This news is about a tool that gives Claude memory across chats by storing and searching your notes locally. RAG, or retrieval-augmented generation, means the AI looks up relevant information before answering, instead of trying to remember everything on its own. Non-experts should care because it shows how AI assistants can be made more useful without sending private notes to a cloud service.

🦞 OpenClaw angle

If you build AI automations, treat local memory as a separate service instead of baking it into your agent prompts. Keep notes in Markdown, store embeddings in a local database, and make retrieval explicit so you can audit what the agent is pulling in. If you need privacy or offline use, prefer offline embedding models and avoid designs that require API keys or external telemetry.

Memex is built around local RAG, or retrieval-augmented generation. In plain terms, that means it looks up relevant notes before responding, so Claude can pull in past context when needed instead of relying only on the current chat.

The project also supports Obsidian-compatible folders out of the box. Users can point Memex at existing note directories with --from and --to, and it can index external folders such as an Obsidian vault. The tool also supports [[Title]] syntax, which links notes together.

A key feature is get_note, which automatically shows what other notes reference a given note. That creates a graph of connected thinking without any manual wiring, according to the project page.

Memex also includes a digest command that prints a summary of everything saved in the last N days, grouped by folder. The developer says this is meant to help users review what Claude captured during the week.

Privacy is a central part of the pitch. According to the project description, Memex uses an embedding model that runs entirely offline with @huggingface/transformers. There are no telemetry calls, no accounts, and no API keys required. Once the model is downloaded, it works without an internet connection.

The project targets developers who want Claude to remember context without putting their notes into a third-party database. It is set up for local use first, with commands and configuration stored under ~/.memex/models/. The source post frames it as a way to give Claude longer-term recall while keeping the underlying notes under the user’s control.

Memex is available now on GitHub, and the project description says it can be started in about two minutes.

Source: HN Show HN ↗

More from OpenClaw News