M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
C

Claude Mem

Community
Persistent cross-session memory for Claude Code and other agents
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #258 of 438
Stars
★ 111.9k Very popular
Runtime
Node.js 20+ · Bun
Credentials
No credential needed
License
Apache-2.0
Last commit
Tools
3
50FMRS · D

claude-mem is a well-documented, open-source memory compression system built primarily for Claude Code. It combines hook-based capture, AI summarization, and a 3-layer MCP retrieval workflow (search/timeline/get_observations) with local SQLite storage and optional cloud sync. Users should be aware of its local HTTP worker service, default full-capture behavior, and an unrelated third-party token associated with the project.

Strongest · Documentation 15/20 Weakest · Security and permissions 5/20

Reliability
6/20
Security and permissions
5/20
Maintenance
14/20
Documentation
15/20
Setup experience
10/20
Why each score
Reliability 6/20
No server manifest was cached, and the README shows no CI workflow or committed test evidence. The system is architecturally complex (a Bun-managed worker service, SQLite, a Chroma vector database, and 7 hook scripts), and its runtime dependencies (Bun, uv) are described as 'auto-installed if missing' — an implicit install step that is itself a common source of cold-start fragility. The README also claims '4 MCP tools' but only describes three (search, timeline, get_observations), an internal inconsistency between the claim and the documented tool set. Per the static calibration rule, reliability cannot exceed 12 without verifiable execution evidence (CI plus committed tests); given the near-total absence of such evidence, the score is further reduced to 6.
Security and permissions 5/20
Several compounding data-boundary and trust risks are visible: (1) the OpenClaw install path is `curl -fsSL ... | bash`, executing an unreviewed remote script directly; (2) the tool 'captures everything your agent does during sessions' and by default routes tool-call content to an unspecified 'AI' provider for compression, with privacy relying only on a manual opt-out `<private>` tag rather than default data minimization; (3) the 'Cloud Sync' feature can, by default use, sync memory data to the third-party domain cmem.ai, with no disclosure in the README of transport encryption, retention policy, or an explicit consent flow; (4) additional runtimes (Bun, uv) are auto-installed, expanding the attack surface; (5) the README ends with promotion of an unrelated cryptocurrency token (CMEM, including an on-chain contract address), which — while not itself a traditional red-line category (malware, credential theft, etc.) — materially undermines confidence in the project's trustworthiness. Given this stack of unmitigated risks and the absence of any least-privilege or confirmation description, the score is set low rather than in the red-line range.
Maintenance 14/20
Clear signs of active maintenance: version 13.4.0, an unambiguous Apache-2.0 license, layered docs covering architecture/configuration/development/troubleshooting, a documented release-branch strategy (main/core-dev/community-edge), official Discord and X presence, and an automated bug-report generator. However, governance gaps remain: no SECURITY.md or dedicated security-response channel is mentioned, and 358 open issues suggest a sizable unresolved backlog. The unrelated cryptocurrency-token endorsement at the README's end is a reasonable basis to question maintainer judgment and project focus. Net score: 14.
Documentation 15/20
Documentation breadth is strong, with layered links covering installation, usage, search tools, cloud sync, context engineering, architecture (hooks, worker service, database, search architecture), configuration, and development, plus README translations in 30+ languages and code examples for MCP tool usage. However, most detailed documentation lives on an external site (docs.claude-mem.ai) that could not be verified in this static review; the README itself discloses no concrete cost, rate limits, AI-provider choice, or data-retention specifics, and provides no copy-pasteable MCP client configuration JSON. Scored 15 — structurally complete but with undisclosed assumptions and verification gaps.
Setup experience 10/20
Multiple install paths are offered (npx script, plugin-marketplace commands, a one-line curl installer for OpenClaw, multi-IDE support), which is broad coverage. However, the dependency chain is heavy: Node ≥20, Bun (auto-installed), uv (auto-installed), and SQLite, with implicit auto-installation of multiple runtimes adding real-world deployment uncertainty and platform-compatibility risk; no standard MCP client (e.g., Claude Desktop) command/args configuration JSON is shown. Per the static calibration rule, setup cannot exceed 15 without verifiable execution evidence; combined with the documented-but-fragile, dependency-heavy install flow, the score is set to 10.

Static review · not runListed 2026-08-06

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesConnects to a database

Best for

  • Developers working on long-running projects with Claude Code (or other supported agents) who want memory to persist across sessions
  • Teams that want automatic summarization of agent tool activity instead of manual note-taking

Not for

  • Privacy-sensitive workflows where agent activity shouldn't be logged/stored by default without manually excluding content via <private> tags
  • One-off, single-session tasks that don't need cross-session context
  • Environments where installing a background worker service (Bun) and a Python package manager (uv) is not permitted

Required permissions

  • Local file system read/write access for session logs, the SQLite database, and configuration (~/.claude-mem)
  • Ability to run and manage a background worker process (Bun) and a Python package manager (uv)
  • Network access for AI-based summarization/compression calls and optional cloud sync to cmem.ai
  • Access to Claude Code (or other agent) lifecycle hooks and tool-use transcripts in order to capture observations

Risks and side effects

  • Captures and stores all tool-use activity from agent sessions by default; sensitive content must be manually excluded using <private> tags
  • Runs a local HTTP worker service with a web viewer, which exposes stored memory data on the local machine/network
  • Optional cloud sync transmits memory data to a third-party service (cmem.ai)
  • Auto-installs additional runtimes (Bun, uv) when missing, widening the trust/supply-chain surface
  • The project is associated with a third-party cryptocurrency token (CMEM) that is not part of the core software — a separate financial-risk consideration, not evidence of malicious code

Setup

Before you start

Runtime：Node.js 20+ · Bun

Other optional settings (1)
CLAUDE_MEM_MODE optional Set in ~/.claude-mem/settings.; controls workflow mode and observation language, e.g. code--zh for Simplified Chinese.

Run npx claude-mem install (use npx claude-mem install --ide opencode for OpenCode, or follow the Antigravity CLI setup guide for Antigravity). Alternatively, install from inside Claude Code with /plugin marketplace add thedotmack/claude-mem followed by /plugin install claude-mem. Restart Claude Code afterward — context from prior sessions then appears automatically in new sessions. Note: npm install -g claude-mem only installs the SDK/library and does NOT register plugin hooks or set up the worker service; the maintainers explicitly recommend the install methods above instead. OpenClaw gateway users can run curl -fsSL https://install.cmem.ai/openclaw.sh | bash for a one-command setup.

Check that it works

Run npx claude-mem install and restart Claude Code, then ask "search my memory from previous sessions"; if the search tool appears in the tool list and returns historical observations, the install worked.

Troubleshooting

  1. Restart Claude Code after installation or after changing CLAUDE_MEM_MODE for settings to take effect
  2. Confirm Node.js >=20 and that Bun and uv are installed (they auto-install if missing)
  3. Verify the worker service is running; the web viewer URL is printed at startup
  4. Check ~/.claude-mem/settings.json for AI model, port, and log level configuration
  5. Describe the issue to Claude to trigger the built-in troubleshoot skill for automatic diagnosis
  6. Generate a detailed bug report with `npm run bug-report` from the plugin marketplace directory

Things to try

Once connected, you can ask your AI assistant things like:

  • Search my memory for notes about the authentication bug
  • Use timeline to show what happened around observation #123
  • Fetch full details for observations #123 and #456
  • Search past sessions for bugfix entries, limit 10

Tools 3

search read-only
Full-text search over the memory index, filterable by type, date, and project; returns a compact result index (~50-100 tokens per result) as step one of the 3-layer retrieval workflow.
timeline read-only
Retrieves chronological context around a specific observation or search result to show what was happening at that point.
get_observations read-only
Fetches full observation details for given IDs (batch multiple IDs for efficiency) — the most detailed, highest-token step of the 3-layer workflow.

Use cases

Automatically restore prior project context when starting a new Claude Code session
Search past coding sessions in natural language to recall earlier decisions or bug fixes
Maintain continuity of knowledge across long-running, multi-session projects

Supported clients

Claude Code
OpenCode
OpenClaw
CodexPartial support
GeminiPartial support
GitHub CopilotPartial support

Listed from the project's documentation, not tested by this site.

Overview

claude-mem is an open-source (Apache-2.0) memory compression system built primarily for Claude Code. It uses 5 lifecycle hooks to automatically capture an agent's tool-use activity during a session, compresses it into semantic summaries with AI, stores it in a local SQLite database, and injects relevant context back into future sessions. A local Bun-managed worker service exposes an HTTP API and web viewer, and a Chroma vector database provides hybrid semantic + keyword search for retrieval. Beyond Claude Code, the project ships install paths for OpenCode and the OpenClaw gateway, and claims compatibility with Codex, Gemini, Hermes, and Copilot, though installation specifics for those clients are not documented.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 64.1k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision 02584c4ac9d9 Data synced 2026-11-04 Read the FMRS scoring method