claude-mem: 108K-Star Open-Source Memory for AI Coding Agents

Every coding agent wakes up with amnesia. You spend an hour teaching it how your checkout flow works, which retry policy the team agreed on, and why the deploy script fails on Tuesdays. Then the session ends, the context window closes, and the next session starts from zero. That gap between what an agent learned and what it remembers is exactly the problem a new open-source project set out to fix, and developers have noticed: claude-mem now carries 108,942 stars and 9,544 forks on GitHub, read live from the GitHub API on November 2, 2026, making it one of the most-starred agent projects on the platform. An independent tracking report published November 1 flagged it as the biggest acceleration jump of the week: its star pace leapt from 254 a day to 882 a day in seven days.

This article explains what claude-mem is, how its memory engine works, what it costs (the free engine plus a $30/month hosted tier), and where the honest catches are.

What claude-mem actually does

claude-mem is a persistent memory compression system for AI coding agents, built by developer Alex Newman (@thedotmack). In plain terms: it watches your agent sessions, captures observations like decisions, bug fixes, and discoveries, turns them into semantic summaries, and stores them in a local database. When a new session starts, the agent can search that history and load the relevant notes instead of re-learning everything.

The project began as a tool built specifically for Claude Code, Anthropic's agentic coding tool, but it has grown into a multi-agent system. According to the project's own description, it works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, and OpenCode, and the README adds install paths for Cursor, T3 Code, Antigravity, OMP, Pi, DeepSeek Harness, and Grok Bot, plus a search skill for Claude Desktop.

How the memory engine works

claude-mem is built from four moving parts, all described in the project's own documentation:

1. Lifecycle hooks

Small commands plug into the agent's events: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and SessionEnd. Every time something notable happens during a session, the hook records an observation. No manual note-taking required.

2. A local worker service

Managed by the Bun JavaScript runtime, it runs an HTTP API with a web viewer so you can browse your memory, review sessions, and search observations from a browser. A "My Work" view shows live requests, progress, and a review queue.

3. SQLite storage with hybrid search

Sessions, observations, and summaries live in a local SQLite database with full-text search, paired with a Chroma vector database for semantic search. Your memory sits on your own machine unless you turn on cloud sync.

4. Progressive disclosure

Memory retrieval follows a three-layer workflow through the project's MCP search tools: search returns a compact index of results, timeline shows chronological context around interesting hits, and get-observations fetches full details only for the IDs you pick. The project claims roughly 10x token savings from filtering before fetching. That figure is the project's own claim with no independent benchmark published, but the design principle is sound: read the summary before loading the details.

Install: one command, then a sign-in choice

Installing claude-mem for most setups takes one command:

npx claude-mem install

The installer sets everything up, then asks you to sign in to claude-mem in your browser with an email magic link. No card is required. Signing in provisions a memory key and starts the 30-day free trial of the hosted observer. Here is the part worth reading carefully: after sign-in you pick your memory provider. The options are the claude-mem observer, your own OpenRouter or Gemini key, or your Anthropic plan. If you would rather skip sign-in entirely, passing an explicit --provider flag, setting CLAUDE_MEM_ONLINE_OPTIN=false, or installing in a CI or non-interactive shell completes the setup with no account interaction.

Other install paths cover specific setups: npx claude-mem install --ide opencode for OpenCode, a one-line installer for OpenClaw gateways, marketplace commands (/plugin marketplace add thedotmack/claude-mem) inside Claude Code, and installers for T3 Code, Antigravity, OMP, Pi, DeepSeek Harness, and Grok Bot. One common mistake to avoid: npm install -g claude-mem installs only the SDK library. It does not register the plugin hooks or set up the worker service, so always install with npx claude-mem install or the plugin commands. The README states this explicitly.

Pricing: free engine, $30/month Pro, and the fallback

The pricing picture is refreshingly simple, which makes it worth laying out exactly:

$30 a month is roughly £22.40 or C$41.92 at November 2, 2026 market rates, verified at publication. The site's own pricing page bills in dollars only, so those conversions are approximations.

Three pieces of pricing fine print matter more than the headline number:

First, what happens after the trial. When the 30-day trial ends, memory automatically falls back to your Anthropic plan unless you subscribe to Pro. In other words, the observation work runs on your own Anthropic usage at Anthropic's rates, which are the published API prices.

Second, the BYO-key option. If you point claude-mem at your own OpenRouter or Gemini key, you pay that provider's rates instead of the Pro subscription. That route is self-managed: you hold the keys and the bills.

Third, the marketing math. The README says the hosted observer means "up to 100% more usage from your plan" because memory generation runs off-plan. That is the project's own claim, not an independently measured figure, and your real savings depend on your workload and provider. The homepage adds one honest line of fine print: CMEM Pro does not change the price or limits of another subscription. You are still bound by whatever your Claude plan allows.

How claude-mem compares to other memory approaches

Persistent agent memory is a crowded idea. Here is how the main options stack up:

The real comparison is not against other products but against habits. If your workflow is a long-running project where the agent rediscovers the same facts every week, claude-mem automates what a disciplined developer would otherwise maintain by hand in memory files. If you open a fresh chat for every one-off task, the hooks add overhead without payoff.

The catches, stated plainly

Your observations flow through a service when you use Pro

The local engine keeps memory in a local SQLite database, which stays on your machine. But the hosted observer, cloud sync, and private MCP link are cloud services by design. The project offers <private> tags to exclude sensitive content from storage, and screenshots are opt-in and off by default, so use both if you work in codebases with secrets.

The default install leans toward Pro

The installer defaults to the hosted memory path with the trial. Staying fully local is possible, but you must opt into it with a provider flag (for Grok Bot, --provider host). If you are cost-sensitive or privacy-strict, choose deliberately instead of clicking through.

Memory quality depends on the observer model

Observations and summaries are AI-generated. Garbage in, garbage out applies: if the observer model misreads a session, the stored note is wrong in a way that looks authoritative later. The project's citation system, which references past observations with IDs through the worker API or the web viewer, helps you trace a note back to its source. Use it before trusting an old memory.

No independent evaluation exists

The 10x token-savings figure, the star counts, and the quality claims are all project-reported or community-observed. The whatstrending.ai acceleration report independently confirms the star velocity, but nothing independently measures memory quality or token savings.

ChatGPT capture has a real limit

The homepage states that MCP connections provide recall, but connecting ChatGPT does not automatically capture ChatGPT conversations. If your workflow lives in ChatGPT, claude-mem recalls synced notes there but does not build them there.

The CMEM token is not the product

The README includes a note that CMEM is a token created by a third party and embraced by the project's creator. It is a community token, unrelated to the tool's pricing. Buy nothing expecting it to lower your bill.

Pro tips and common mistakes

Install with npx claude-mem install, not npm install -g. The npm package is the SDK only and will leave you wondering why nothing is hooked up.

Restart your IDE after installing. The hooks register on the next launch, so memory will not appear until you do.

Decide your provider up front. Trial-to-Pro-to-fallback is a fine default for testing, but a team rolling this out should choose the local engine, BYO key, or Pro deliberately rather than discovering the fallback in month two.

Tag sensitive content with <private>. API keys, customer data, and credentials that pass through sessions can otherwise land in observations.

Check the web viewer weekly at first. The memory loop works best when someone occasionally prunes or corrects notes, the same way a wiki needs gardeners.

Who should try it, and who should skip it

claude-mem fits developers who run the same coding agent across long-lived projects and are tired of re-teaching it. It fits teams already using multiple coding agents, since one memory store can serve Claude Code, Codex, OpenCode, and the rest rather than fragmenting per tool. It fits privacy-minded developers too, as long as they stay on the local engine and skip cloud sync.

Skip it if your agent work is mostly short, one-off sessions. Skip the Pro tier if a disciplined memory file and native session compaction already cover your needs. And skip any assumption that 108,942 stars mean the memory is accurate: star counts measure attention, not reliability.

Is claude-mem really free?

Yes. The local engine is open source under Apache 2.0 and costs nothing. CMEM Pro, the hosted observation and cloud-sync service, is $30 a month after a 30-day free trial. The project's own FAQ answers this directly.

Does it work with coding agents other than Claude Code?

Yes. Although it started as a Claude Code plugin, the project now supports Codex, OpenCode, OpenClaw, Cursor, Copilot, Gemini, Hermes, T3 Code, Antigravity, OMP, Pi, DeepSeek Harness, and Grok Bot, each with its own install path in the README.

Where is my memory stored?

Locally, in a SQLite database with full-text search, unless you enable cloud sync through CMEM Pro. You can exclude content with <private> tags, and screenshots stay off unless you opt in.

What happens after the 30-day Pro trial ends?

Memory falls back to your Anthropic plan automatically unless you subscribe to Pro. You can also point the observer at your own OpenRouter or Gemini key and pay that provider instead.

Does it capture my ChatGPT conversations?

No. Connecting ChatGPT gives recall of synced memory through an MCP connection, but ChatGPT conversations themselves are not automatically captured. Capture depends on the tool and integration you install.

How do I verify the star count and pricing myself?

The star count is public on the GitHub repo page, and pricing is on cmem.ai. Both were read live during the writing of this article on November 2, 2026.

The short version: claude-mem attacks the most annoying property of coding agents, the way they forget everything between sessions, with an open-source engine you can self-host for free. The $30/month Pro tier buys convenience, cloud sync, and off-plan observation, not access to the tool itself. Install it, point it at your provider of choice, tag your secrets, and let the agent finally remember what it decided last Tuesday.