What it is
claude-mem gives Claude Code and other coding agents memory across sessions. Six lifecycle hooks (session start, prompt submit, before and after each tool call, stop and session end) capture what the agent did; a local worker managed by Bun compresses those observations into summaries with a language model and stores them in SQLite, with Chroma for semantic search; relevant context is injected at the start of the next session.
Recall runs through MCP tools in three steps (search, timeline, get_observations) so the agent fetches full detail only for the records it needs; v15 added a faster, shallower dumb_search. A local web viewer shows what has been stored, and text inside <private> tags is excluded. Besides Claude Code it installs into Codex, OpenCode, T3 Code, Antigravity CLI, OpenClaw, Hermes Agent, Pi and other harnesses.
The author, Alex Newman, also sells a hosted memory service, CMEM, which the installer offers on a 30-day trial.
Our review
claude-mem does what it promises, and it has also become the front door to a paid service. The repository has about 102,400 GitHub stars and shipped v15.0.1 on 20 October 2026. Run the installer and, once it has finished, it opens a browser for an email sign-in that starts a 30-day trial of the hosted observer; when the trial ends, memory falls back to your Anthropic plan unless you subscribe. Passing --provider or setting CLAUDE_MEM_ONLINE_OPTIN=false skips the account step entirely.
The local design is sound. Observations live in ~/.claude-mem, in SQLite with full-text search plus a Chroma vector index, and recall is index-first: the README puts a search result at roughly 50 to 100 tokens and a full observation at 500 to 1,000, so the agent pays for detail only when it asks. The worker listens on 127.0.0.1, on port 37700 plus your user ID modulo 100, and the configuration docs state plainly that it has no request authentication. Binding it to 0.0.0.0 without a token, which the docs call dangerous, would expose your project history to the local network.
The running cost is compression. Every session is summarised by a model, Claude Haiku 4.5 by default when the provider is your Claude plan. GitHub issue #2643, filed in May 2026, describes 649 million tokens spent across about 530 observer sessions before the user noticed; it was closed two days later, and v15 adds fallback between an API key and a subscription. The release pace is its own risk: v15.0.0 arrived three days after v14.0.1 and removed the old mem_search and flash_search calls.
The hosted tiers run from Pro at $30 a month to Max at $100 with metered overages, and the included allowances appear only at checkout. Cloud sync means CMEM stores your observations, and the pricing page says nothing about retention or training use. The README also endorses a third-party crypto token, also called CMEM, on the Base chain, which has nothing to do with the software. Against a well-kept CLAUDE.md, which costs nothing and runs no daemon, claude-mem only pays its way on a codebase you return to for weeks.
Where it earns its place
A solo developer maintaining one SaaS codebase for months with Claude Code
Stop re-explaining the architecture at the start of every session
Install with --provider pointing at the Claude plan and leave the Haiku default; past decisions and bug fixes come back at session start. The software is free, but compression draws on the plan, so check usage in the first week given issue #2643.
A contractor moving between Claude Code and Codex on client repositories
Keep one project memory across both agents without a cloud account
Run local-only and set CLAUDE_MEM_SESSION_START_INCLUDE_ALL_SOURCES so both harnesses see the same observations. Wrap client credentials in <private> tags; nothing leaves the machine except the text sent to the compression model.
An engineering lead on a ten-person product team
Share what one developer's agent learned with the rest of the team
TeamBrain is a pilot priced on a call, and it relies on cloud sync, so CMEM holds the shared observations. Retention and training terms are not published on the pricing page, which matters before client code is involved.
Before you commit
- Setup
- `npx claude-mem install` (Node.js 20+; Bun and uv are installed if missing) or the Claude Code plugin marketplace. The installer asks for an email sign-in unless you pass --provider or set CLAUDE_MEM_ONLINE_OPTIN=false.
- Integrations
- Claude Code, Codex, OpenCode 1.3.4+, T3 Code, Antigravity CLI, OpenClaw, Hermes Agent, Pi and others; MCP tools search, timeline, get_observations and dumb_search; compression via your Anthropic plan or key, OpenRouter, Gemini or the hosted observer.
- Data and privacy
- Stored locally in ~/.claude-mem (SQLite plus Chroma); session text is sent to whichever compression provider you choose. The worker binds to 127.0.0.1 with no request authentication. Cloud sync stores observations with CMEM; retention and training use are not published on the pricing page.
- Cost at scale
- Free to self-run, but every session costs compression tokens (Haiku 4.5 by default). Hosted plans are $30, $70 or $100 per month with allowances shown only at checkout; TeamBrain pricing is not announced.
- Limits
- Six hooks run on every session; 132 open issues; major versions arrive days apart and v15 removed the old search API, so refresh MCP connections after upgrading.
What's good
- Captures memory automatically through lifecycle hooks
- Local SQLite and Chroma store with a viewer to audit it
- Index-first recall, about 50 to 100 tokens a result
- Installs into Claude Code, Codex, OpenCode and other harnesses
What's not
- Every session is summarised by a model you pay for
- Installer defaults to a sign-in and hosted trial
- Local worker API has no request authentication
- Breaking changes between major versions days apart
Best for
Developers on long-running projects who want the agent to remember past sessions.
Alternatives
claude-mem: common questions
- Is claude-mem worth it?
- Real cross-session memory for Claude Code, wrapped in an installer that steers you towards a hosted subscription and a compression bill you have to watch. We rate it 3.7 out of 5.
- How much does claude-mem cost?
- Open source (Apache-2.0). Compression runs on your Anthropic plan or your own key. Hosted CMEM plans: Pro $30/month, Heavy $70/month, Max $100/month with overages; TeamBrain pricing not announced.
- What are the downsides of claude-mem?
- Every session is summarised by a model you pay for; installer defaults to a sign-in and hosted trial; local worker API has no request authentication; breaking changes between major versions days apart.
- Who is claude-mem best for?
- Developers on long-running projects who want the agent to remember past sessions.
- How much does claude-mem cost for a team or in production?
- Free to self-run, but every session costs compression tokens (Haiku 4.5 by default). Hosted plans are $30, $70 or $100 per month with allowances shown only at checkout; TeamBrain pricing is not announced.
- Is claude-mem safe for company data?
- Stored locally in ~/.claude-mem (SQLite plus Chroma); session text is sent to whichever compression provider you choose. The worker binds to 127.0.0.1 with no request authentication. Cloud sync stores observations with CMEM; retention and training use are not published on the pricing page.
- What are the best claude-mem alternatives?
- The closest alternatives we have reviewed are claude-obsidian, Graphify and Everything Claude Code (ECC).
Answers checked Oct 23, 2026.