← All reviews

claude-mem keeps notes between coding sessions

claude-mem adds memory to Claude Code by recording activity, using Claude Haiku to summarize it and bringing notes into later sessions.

A slate-blue pigeon in a teal apron stands on a book press handle above folded pages, while the orange Claude mascot waits beside a small notebook in a crowded bookbinding shop.
claude-mem saves Claude the recap, though two folders with one name share the notebook.

Starting a fresh coding chat often means explaining decisions you already made. claude-mem adds memory to Claude Code, Anthropic’s coding agent. Its background note-writer can use Claude Haiku to record what happened while you worked.

Tested: commit c4a9f5725f420bf9c6e866cdd3b94bd36498af09 on November 12, 2026, on Linux with Claude Code and Claude Haiku. Workflow: record a service’s development history, start fresh sessions that must reuse its decisions, and check two repositories that share a folder name. Result: claude-mem supplied earlier coding decisions to a fresh session, though built-in memory already held the same decisions. Two independent repositories with root folders named alike shared automatically captured notes. Verdict: adopt with care, choosing distinct root-folder names or Git-remote naming before recording work.

Set up claude-mem with your coding agent

Paste this into your coding agent:

Install the latest claude-mem from https://github.com/thedotmack/claude-mem.
Before recording work, name each project's memory by its Git remote.
Turn off anonymous usage analytics.
Show that captured project notes appear in a fresh session.

claude-mem records activity alongside built-in memory

Claude Code’s built-in auto memory lets the coding agent choose which notes to save. claude-mem captures tool activity through hooks (callbacks when the agent acts) and sends it to a separate note-writer. New sessions automatically receive recent notes, and the agent gets 19 extra memory tools, including search.

For a service that sends requests and retries failures, a fresh coding session started with this note (excerpt; retry delays are in milliseconds):

### Nov 12, 2026
3 " ⚖ Provisional follow-up ideas require user confirmation before implementation
4 " ⚖ Delivery retry and dedupe rules decided, superseding provisional [100, 200] ms and payload-hash ideas
...
**Learned**: The decided new-client delivery rules are retry_delays_ms [900, 2700, 8100] for status 429 and 500-599 only, dedupe by event_id alone, and no retries for duplicates.

The dated list marks decisions with ⚖ and records which ideas they replaced. A summary of what was investigated, learned and completed follows, and the whole note ran to about 800 words. The session then made the correct change, though it had also read Claude Code’s own saved note holding the same rules.

You can pause claude-mem’s note-writing when usage rises

In our service test, recording the history and finishing one follow-up change processed about 1.1 million tokens with claude-mem, against about 380,000 with Claude Code’s built-in memory alone. Most of the difference was Haiku writing notes while the history was recorded:

Step Built-in memory only With claude-mem
Recording the history (coding agent) 349,181 349,181
Notes on the recording (Haiku) 0 628,732
Follow-up change (coding agent) 31,220 58,997
Notes on the follow-up (Haiku) 0 58,323
Total 380,401 1,095,233

The counts include cached input at face value, so they show where the work went, not a bill. The follow-up session with claude-mem also carried the startup note and its search tools. Notes on one small invoice change took Haiku about 11,600 tokens. If you use your Claude subscription for summaries, that note-writing shares your plan allowance.

To pause new summaries, set CLAUDE_MEM_OBSERVER_ENABLED to "false" in ~/.claude-mem/settings.json. Existing memory stays searchable. Claude Code’s /usage command shows plan usage.

Coding activity goes to Haiku; claude-mem stores notes locally

Recorded tool inputs and outputs can include source code. A background service on your computer stores the notes and permits access by other local programs; use a trusted development computer. A security report found that any account on the same machine could read every note and change claude-mem’s settings through that service. Version 15.18.3, released after our test, adds an optional CLAUDE_MEM_LOCAL_API_TOKEN setting that makes local programs present a token. On a machine with other users, set it in ~/.claude-mem/settings.json and restart claude-mem’s background service.

Anonymous usage analytics are on by default; npx claude-mem telemetry disable turns them off.

Same-name project folders share memory by default

In our test, two separate repositories with the same root-folder name shared memory: a fresh session in one received the other’s retry and currency policy. An Android developer reported independent repositories sharing a memory label. claude-mem uses the root-folder name to choose memory, ignoring the parent directory.

Give independent repositories distinct root-folder names before capture, or name memory by Git remote in ~/.claude-mem/settings.json:

{
  "CLAUDE_MEM_PROJECT_NAME_SOURCE": "git-remote"
}

With Git-remote naming, repositories with different owner/repository names keep automatically captured notes separate. Local-only repositories still use folder names. Enabling Git-remote naming keeps old mixed notes in startup context; use reviewed project files for those decisions.

Users report heavy background usage on older versions

On version 9.0.12, a developer added up four days of Claude Code logs: note-writing used 245.5 million tokens, against 348.4 million for all their own work. The maintainer replied on November 10 that the note-writer can now be turned off, folder CLAUDE.md files are off by default and another provider can write the notes instead of your Claude plan. Another developer gave claude-mem its own API key and a cheaper Flash model; after about a month it cost $65.82, against $102.94 for their coding model.

In the project’s feedback thread, one developer says combining claude-mem with Markdown progress files makes switching branches and recovering project knowledge easier. A Reddit user says their $200-plan allowance lasts about five to six days across two or three projects, with Codex taking some of the work and progress notes kept in Obsidian.

In April, a developer traced stalled unattended runs to claude-mem cutting file reads short. The current version blocks whole-file reads of code files of 32 KB or more that already have notes and points Claude to claude-mem’s outline tools. If long unattended runs stall, set CLAUDE_MEM_FILE_READ_GATE_ENABLED to "false".

Adopt claude-mem alongside reviewed project notes

Adopt claude-mem with care for automatic coding notes if you choose distinct root-folder names or Git-remote naming before recording work. Claude Code’s built-in auto memory offers a simpler way to save selected decisions. basic-memory offers editable Markdown notes shared with agents; claude-supermemory offers automatic capture through a hosted service. If strict project separation is essential and you want fewer settings to manage, keep decisions in explicit project notes.