Claude-Mem

Claude-Mem is a Claude Code plugin from Alex Newman (GitHub: [thedotmack](https://github.com/thedotmack)). Its purpose is persistent working context: what an agent did during a session, which decisions were made, which errors occurred, and which next steps remain should not live only in a short-lived chat transcript. The official repository now calls the product “Grok Mem”, while its Claude Code plugin, marketplace entry, and npm package continue to use claude-mem. This catalog entry therefore uses the installable Claude-Mem name and makes the rename explicit. Official documentation is available at [docs.claude-mem.ai](https://docs.claude-mem.ai/), and complete source is in [thedotmack/claude-mem](https://github.com/thedotmack/claude-mem).

What the plugin does in Claude Code

Per the official README, Claude-Mem registers lifecycle hooks for session start, prompt submission, tool use, stop, and session end. A local worker turns those events into observations and summaries; a SQLite database stores sessions, observations, and summaries. The provider also documents a Chroma vector store for search and a mem-search skill. The intended design is not to inject every old transcript into context without reduction. Search begins with a compact index, can show chronological neighbourhood when useful, and retrieves fuller observations only for selected IDs. That is meant to control context consumption while retaining the ability to find past work.

In practical terms, this can help with longer development work: after a restart, a new Claude Code chat may find prior decisions, files already inspected, or the state of a bug instead of exploring the entire codebase again. Search can also answer project-specific questions about an earlier migration, a discarded approach, or outstanding follow-up work. It supplements a model’s own memory; it is not a guarantee that every relevant item is captured, summarized correctly, or injected for every request. Critical architectural decisions, requirements, and approvals still belong in versioned documentation, tickets, and code review.

Installation through the Claude Code marketplace

The primary source confirms Claude-Mem is a valid plugin in its own marketplace: .claude-plugin/marketplace.json lists claude-mem, sourced from ./plugin, as a persistent context system for Claude Code. In a Claude Code session, first add the marketplace:

/plugin marketplace add thedotmack/claude-mem

Then install this entry:

/plugin install claude-mem

Restart Claude Code afterward. The provider also documents npx claude-mem install as an alternative; it configures dependencies, plugin hooks, and the worker. npm install -g claude-mem is not equivalent: per the official README, it installs only the SDK or library, not the hooks and worker service. The source lists Node.js 20 or later, current Claude Code with plugin support, and Bun and uv as runtime requirements that the installer can add automatically. In an enterprise network, teams should first confirm that package downloads and their selected model or hosting path are allowed.

Data flow, privacy, and security boundaries

Claude-Mem explicitly observes prompts and tool use so that later context can be created. That makes deliberate scope especially important in confidential repositories. The official README documents <private> tags for excluding sensitive content from storage. This is useful, but it is not permission to let credentials, keys, personal data, or production secrets casually appear in prompts and tool output. Repository safeguards, secret scanning, least-privilege access, and human review still matter.

The documentation describes local SQLite and Chroma components, a local worker with a web viewer, and optional cloud sync to cmem.ai. Processing is not determined by the plugin alone, though: during setup, the provider says a hosted observer, a personal OpenRouter or Gemini key, or an Anthropic plan can be selected as the memory or model path. When a hosted service or external model provider is involved, observations or content needed for summarization can reach that service; its retention, access, and compliance terms must be reviewed before production use. Enable cloud sync only after a deliberate approval. Local storage also requires protection: do not expose the working directory, database files, or web viewer broadly to other users or the network.

License, popularity, and fit

The repository is licensed Apache-2.0. The 93,367 GitHub stars were recorded directly through the GitHub API on 2026-09-07; they describe the full repository, not the Claude Code plugin alone. Stars are a point-in-time popularity signal, not proof of quality, privacy, or security. The source describes optional provider and hosting paths, so this entry does not state fixed prices. Before adoption, teams should review the official plans and privacy terms for the route they actually select.

Claude-Mem is most suitable for recurring agent work in long-lived codebases where sessions are frequently interrupted and retaining a traceable thread of work is useful. It is less suitable for short one-off tasks, strictly isolated environments without an approved storage path, or teams that do not allow automated observation of prompts and tool output. A sound first trial is a non-sensitive test project: complete the marketplace installation, finish a small task, restart Claude Code, ask a focused question about that stored work, then inspect data placement, viewer access, and excluded content.

/plugin install claude-mem