mcp-memex 0.1.5

Persistent session memory for Claude Code — local SQLite, no LLMs, no network

memex

Every Claude Code session starts without knowing what you built last time — which files you changed, what decisions you made, what broke. You re-explain. Claude re-reads. You start over.

memex keeps a structured log of every session. Claude reads it at the start of the next one.

No cloud. No LLM extraction. No cost per save. Just local SQLite and two commands to install.

Install

pip install mcp-memex
memex install

Restart Claude Code. That's it.

What gets saved

Each entry has structure — not a blob of text:

mem_save(
    task="Replaced JWT auth with session cookies",
    files=["src/auth.py", "src/middleware.py"],
    decisions=["Session cookies over JWT — simpler, no token refresh needed"],
    warnings=["Redis required — app won't start without REDIS_URL set"],
    tags=["auth", "sessions"]
)

Claude calls mem_save when it finishes something meaningful. You can also just tell it: "save what we just did."

What Claude sees at session start

=== memex: session context (db: my_project.db) ===

--- Recent (2) ---
[1] #4 — 2026-06-14T10:32
  Task      : Replaced JWT auth with session cookies
  Files     : src/auth.py, src/middleware.py
  Decision  : Session cookies over JWT — simpler, no token refresh needed
  Warning   : Redis required — app won't start without REDIS_URL set
  Tags      : auth, sessions

[2] #3 — 2026-06-13T15:10
  Task      : Added rate limiting to /api/login
  Files     : src/middleware.py
  Warning   : Rate limiter is in-memory per process — resets on restart
  Tags      : auth, rate-limiting

Claude calls mem_load automatically at the start of every session. It returns the most recent entries plus any entries that match what you're working on — by keyword and by file path.

Browse from the terminal

You don't need to be inside Claude to look at your history:

memex status                # entry count, DB size, top tags
memex list                  # recent entries for this project
memex list --tag auth       # filter by tag
memex search "rate limit"   # full-text search

Export to markdown

Turn your history into portable, human-browsable notes — one .md file per entry, with [[wikilinks]] connecting related entries and files, plus tags as #hashtags. Works in Obsidian or any tool that understands [[links]]:

memex export ./my-project-notes/

Entries that share a tag or a touched file are cross-linked under a Related section, so you can navigate your project history as a graph.

Six MCP tools

Tool	What it does
mem_load	Called at session start — returns recent + relevant entries
mem_save	Saves a structured entry after meaningful work
mem_update	Patches specific fields of an existing entry (keeps id and timestamp)
mem_search	Full-text search across all entries
mem_list	Lists entries, optionally filtered by tag
mem_delete	Removes a stale entry by id

Why not just CLAUDE.md?

CLAUDE.md is for static project documentation — architecture, conventions, how to run tests. It doesn't change much and it isn't session-aware.

memex captures what's changing session to session: what you built yesterday, the decision you made this morning, the warning you discovered an hour ago. It's the difference between "here's the project" and "here's what happened last time."

memex install writes Claude's instructions into .claude/memex.md — a file Claude Code loads automatically without touching your project's own CLAUDE.md.

Memory is scoped per project

Each project gets its own SQLite database at ~/.memex/<project>.db based on the working directory. Sessions from different projects never mix.

Configuration

Set these in the MCP env block in ~/.claude.json if you need to override defaults:

Variable	Default	Description
MEMEX_DIR	~/.memex	Where DBs are stored
MEMEX_GLOBAL	0	Set to 1 to share one DB across all projects
MEMEX_RECENT	5	Max recent entries loaded per session
MEMEX_MATCHED	5	Max FTS-matched entries loaded per session

Uninstall

memex remove
pip uninstall mcp-memex

Memory DBs are kept at ~/.memex/ — delete that directory manually if you want to wipe everything.

Requirements

• Python 3.10+
• mcp package (installed automatically)
• SQLite with FTS5 (standard since Python 3.8)
• Claude Code CLI

License

MIT