claude-mem: persistent memory for Claude Code

• Searches past sessions for what you did and decided

  Looks up stored notes in three steps: search, then a timeline around a hit, then the full notes. Use it to ask whether something was already solved.

  Try: “Did we already solve this? How did we fix the auth bug last time?”

  Read · mem-search

• Writes a story of a project from its memory

  Reads a project's whole stored history and writes a narrative report of how it developed.

  Try: “Give me a timeline report for this project.”

  Write · timeline-report

• Writes a week-by-week digest of a project

  Walks the stored timeline one week at a time and writes a short narrative for each.

  Write · weekly-digests

• Writes a HANDOFF.md so a new session can continue

  Captures the goal, current state, files touched, failed attempts and next steps in one file.

  Try: “Write a handoff so a new session can pick this up.”

  Write · handoff

• Explains what is captured and where it lives

  Describes how claude-mem records tool calls, when notes are injected and where the data sits on disk.

  Read · how-it-works

• Builds a question-answering knowledge base from notes

  Turns stored notes into a knowledge base you can query in plain language.

  Write · knowledge-agent

• Searches code structure with fewer tokens

  Uses tree-sitter parsing to outline and unfold code instead of reading whole files.

  Files · smart-explore

• Sets up, checks or restores database snapshots

  Manages local snapshots of the claude-mem database.

  Runs code · backup-setup

• Sets up or checks Cloud Sync with cmem.ai Pro

  Walks through turning on sync, which uploads notes and full prompt text to your cmem.ai account.

  Write · cloud-sync

• Creates custom memory modes

  Interactively builds a mode with its own note types and tags, and can restart the worker.

  Runs code · mode-creator

• Reads every source file to prime Claude

  Reads each file in full so the session starts with the whole codebase in view.

  Files · learn-codebase

• Writes a phased implementation plan

  Drafts a plan with documentation lookups before any code is changed.

  Write · make-plan

• Runs a phased plan with subagents

  Executes a written plan step by step using subagents.

  Runs code · do

• Automates a coding task end to end

  Chains planning, doing and watching a pull request.

  Runs code · build

• Watches a pull request until it can merge

  Follows a PR or review cycle and reports when it is ready.

  Network · babysit

• Automates a release

  Bumps the version and checks the shipped release.

  Runs code · release

• Bumps the version of a Claude Code plugin

  Runs a semantic versioning and release flow for plugins.

  Runs code · version-bump

• Reports agent spending for a period

  Builds a cost report, by default for the last 7 full days.

  Read · agent-cost-report

• Searches Screenpipe history and saves picks to memory

  Queries screen and audio history recorded by Screenpipe and can store selected context.

  Write · screenpipe

• The rest are workflow helpers that mostly do not use memory

  standup, pathfinder, design-is, oh-my-issues, owner-goals, request-progress, childhood, ccs-align, wowerpoint, what-the and build-and-release. See the plugin's skills folder for each.

  Read · 11 more bundled skills

• Local search server that reads the worker's notes

  Runs on your machine over stdio and exposes search, timeline, get_observations, smart_search, outline, unfold, save_memory and corpus tools. The exact set varies by setup.

  Try: “Search my memory for the database migration decision.”

  Write · mcp-search (MCP server)

• Used by the assistant behind the scenes

  Starts the background worker

  Launches the local worker (HTTP API, web viewer, SQLite) if it is not running.

  Runs code · SessionStart: claude-mem start

• Loads past notes into the new session

  Reads stored notes from the local database and adds them to the session.

  Read · SessionStart: context

• Installs the claude-mem launcher

  Compiles the launcher into ~/.local/bin if missing or old, and may append to your shell PATH.

  Runs code · SessionStart: ensure-launcher

• Checks plugin dependencies on init

  Runs only with claude --init or --maintenance and runs bun install if needed.

  Runs code · Setup: version-check

• Saves your prompt and opens the session

  Strips private tags from the prompt, records it and opens the session in the worker.

  Write · UserPromptSubmit: session-init

• Adds file history before reads

  Adds a file's memory timeline as context and can block whole-file reads of large code files.

  Read · PreToolUse: file-context

• Records each tool call

  Sends every tool input and output to the worker, which queues it for the note-taking model.

  Write · PostToolUse: observation

• Records failed tool calls

  Same capture as PostToolUse for calls that failed.

  Write · PostToolUseFailure: observation

• Stores an end-of-turn summary

  Asks the note-taking model to summarize the turn and saves it.

  Write · Stop: summarize

• Records subagent completion

  Marks a subagent as finished in the worker.

  Write · SubagentStop: subagent-stop

• Marks the session complete

  Closes the session in the worker.

  Write · SessionEnd: session-end

Watch out for

• The npx installer may upload recent prompts to cmem.ai

  An open issue reports the interactive npx setup page sends your last 30 days of prompts to cmem.ai before asking. A consent fix is proposed but not merged, and installing from the /plugin marketplace is not described in that issue. github.com

• Hooks record every prompt and tool call

  UserPromptSubmit saves your prompt, and PostToolUse and PostToolUseFailure send each tool input and output to the local worker. Wrap sensitive text in private tags or exclude projects in settings. github.com

• It can block large file reads

  The PreToolUse hook (file-context) adds a file history before Read, Bash, Grep and Glob, and can block a whole-file Read of a large code file that has history. Turn this off in settings if Claude seems to skip files. github.com

• Captured data goes to a model you choose

  Tool events are sent to the note-taking model: Claude through your login by default, cmem.ai if you pick CMEM Pro in the installer, or your own Gemini or OpenRouter key. Check which one the installer picked. docs.claude-mem.ai

• The note-taker uses your Claude plan

  The default provider runs on your Claude Code login with a Haiku model, and users report it using a large share of their limits. Point it at an API key or another provider to stop that. github.com

• The Stop hook asks a model to summarize each turn

  On Stop it runs claude-mem to request an end-of-turn summary and store it; SubagentStop and SessionEnd only mark work done. Set the summary-skip option to turn summaries off. github.com

• SessionStart hooks start a service and edit your PATH

  On session start it launches a local background worker, injects stored notes, and may compile a launcher into ~/.local/bin and append to your shell PATH. The Setup hook can run bun install on claude --init. github.com

• The local worker has no sign-in

  It listens on 127.0.0.1 without authentication. Binding it to 0.0.0.0 or forwarding its port exposes your settings, including provider API keys, to others. docs.claude-mem.ai

• Telemetry is on by default

  Anonymous usage counts go to PostHog, with no prompts or paths per the docs. Turn it off with the telemetry disable command or the DO_NOT_TRACK variable. docs.claude-mem.ai

• Cloud Sync uploads full prompts when you turn it on

  It is opt-in, and the docs say it uploads observation notes and full prompt text to your cmem.ai account. Leave it off if content must stay on your machine. docs.claude-mem.ai

• It installs Bun and uv on first run

  Missing runtimes are downloaded for you, and a Chroma vector index runs through uvx for search. docs.claude-mem.ai

Security advisories

• CCursorSupported
• WWindsurfSupported
• CClaude DesktopSupported
• CClaude CodeHas issues
• CCodexHas issues
• OOpenCodeHas issues
• GGemini CLINot checked

Tap an app to see how to set it up and any known problems.

In their words