claude-mem
A Claude Code plugin that quietly takes notes on your coding sessions and hands them back next time.
by Alex Newman (thedotmack) · Built this? Edit this page on GitHub
In short
While you work in Claude Code, a second, hidden AI watches each step (every file read, edit and command) and writes short notes about what was built, fixed or learned. After each reply it also writes a summary: what you asked, what was found, what's next. When you start again, the newest notes are laid out at the top of the chat, and Claude can search the rest by keywords or meaning, then open just the notes it needs.
Who it’s for
Developers using Claude Code and other coding agents who want the next session to pick up where the last one left off, without writing notes themselves. It installs with one command; a paid hosted plan adds sync across machines.
What it remembers
What happened while coding: bugs fixed, features built, decisions and discoveries, the files each one touched, your prompts, and a summary of each exchange.
- Memory lives
- Your machine (cloud sync optional)
- API keys
- None on your Claude plan; or a CMEM Pro, Gemini or OpenRouter key
- Works with
- Claude Code, Codex, Cursor, OpenCode, OpenClaw, MCP tools
- License
- Open sourceApache-2.0
- Price
- Free; hosted Pro $20–30/mo
Under the hood
- Who writes memories
- A separate “observer” AI (Claude Haiku by default) watches each tool step and writes notes, plus a summary after each reply. The AI you talk to doesn't write them.
- Where they're kept
- A SQLite database on your machine, with a local Chroma index for meaning-based search. Optional sync to their cloud.
- How they're found
- The newest 50 notes and recent summaries are shown at the start. After that, Claude searches by keywords or meaning, looks at what happened around a note, and opens the ones it needs.
- Extra AI calls
- Yes: observer calls as you work, plus a summary each reply. They run on your Claude plan, your own key, or their hosted observer.
How it works
- 1You workClaude reads, edits and runs commands as usual.
- 2A hidden observerAnother AI watches each step and writes short, labelled notes.
- 3Sum it upAfter each reply: what was asked, learned, done, and what's next.
- 4StoreSaved in SQLite on your machine, searchable by keywords and meaning.
- 5Hand them backNext session, the newest notes come first; Claude searches for more.
What stands out
- Fully automatic: there's nothing to remember to save. A silent observer turns raw tool activity into tidy, labelled notes as you work.
- Shows a cheap index first and fetches full detail only when needed, so memory doesn't eat up the conversation.
- Notes tied to files: when Claude opens a file it has worked on before, that file's history comes up with it.
- Huge reach: one of the most popular memory tools on GitHub, with notes in about 30 languages and plugins for many coding agents.
What it doesn’t do
- Notes are written once and kept as they are. Nothing merges, fades or corrects them, so an outdated note sits beside the newer one.
- It remembers the work, not you: no beliefs about people or preferences, and no sense of self for the AI.
- The hidden note-taker is an extra AI working alongside every session, on your Claude plan by default or on their hosted service.
- Anonymous usage stats are on by default (one command turns them off), and cloud sync, if you turn it on, uploads your notes and prompts.
Measured against human memory
0 built, 5 partly, 6 not of the Field Guide’s 11 mechanisms, and 1 in development
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
The observer is told what's worth recording and to skip routine steps, and it labels each note (bugfix, decision…). There's no importance score. source
Emotion
Gives memories written with feeling more weight.
No emotional weighting. Notes are sorted by the kind of work, not by feeling.
Storage
Forgetting
Lets unused memories fade over time on its own.
Nothing fades note by note, but meaning-based search only looks back 90 days by default, and the start of a session shows just the newest notes. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
Exact copies within a session are dropped, but a newer note that contradicts an older one is simply stored beside it. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
Being found doesn't strengthen a note. A usage counter exists in the database, but nothing updates it yet. source
Their open fix plan (issue #3609) says the counter will go up each time a note is recalled.
Association
Links memories, so recalling one brings its neighbours.
Notes are tied to their files and their moment in a session, and a timeline tool brings up what happened just before and after one. Notes aren't linked to each other. source
Prospective
Brings a memory back at the right time without being asked.
When Claude opens a file it has notes on, that file's history is added on its own. There are no reminders for a set time. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
Nothing reworks notes between sessions. A summary is written after each reply, but old notes stay as first written.
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
Nothing corrects an old note when it's contradicted; old and new sit side by side.
Gist
Turns many specific moments into general knowledge.
Each reply's summary records what was learned, and on request a “knowledge agent” answers questions from hundreds of notes at once. No lasting general lessons are written. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
No beliefs about people or projects. An optional CLAUDE.md per folder lists recent activity, not conclusions.
Not every system is trying to be a human memory, so a hollow dot isn’t a flaw. It just says what the system doesn’t do. What the mechanisms are.
Compare claude-mem with the others, side by sideQuestions
- Who is claude-mem for?
Developers using Claude Code and other coding agents who want the next session to pick up where the last one left off, without writing notes themselves. It installs with one command; a paid hosted plan adds sync across machines.
- Where does claude-mem keep its memories?
Your machine (cloud sync optional). A SQLite database on your machine, with a local Chroma index for meaning-based search. Optional sync to their cloud.
- Does claude-mem forget?
Forgetting: a little. Nothing fades note by note, but meaning-based search only looks back 90 days by default, and the start of a session shows just the newest notes. source
- Does claude-mem correct a memory when it turns out to be wrong?
Reconsolidation: not built. Nothing corrects an old note when it's contradicted; old and new sit side by side.
- Is claude-mem open source?
Yes, it's open source. License: Apache-2.0. Code: github.com/thedotmack/claude-mem.
- Does claude-mem work with Claude Code?
Yes. Works with: Claude Code, Codex, Cursor, OpenCode, OpenClaw, MCP tools.
- Does claude-mem need an API key?
API keys: None on your Claude plan; or a CMEM Pro, Gemini or OpenRouter key.
- How much does claude-mem cost?
Price: Free; hosted Pro $20–30/mo.