Delv
Official (Anthropic)Active· 12d2,802t★ 4.6by Anthropic

Memory

Persistent memory graph for Claude. Survives across conversations - useful for personal assistants that need to remember facts about you and your projects.

View repo
A
Safety & Trust

Delv Safety Grade: A

Score 85/100 · assessed 2026-04-22

Maintainer95
Permissions70
Supply chain90
Transparency85
Incidents100

Anthropic's official memory server provides persistent knowledge graph storage across Claude conversations. It's distributed via npm with proper versioning and lives in the official MCP servers monorepo. The maintainer score is excellent given Anthropic's backing. Permissions are moderately scoped: it writes structured data to local filesystem (a SQLite database) and reads it back, but doesn't execute code or access network. Supply chain is solid through npm registry. Transparency is good with open source code and documentation, though changelog detail could be fuller. No known security incidents. The main consideration is that it stores potentially sensitive personal information locally in plaintext SQLite, so filesystem access by other processes could expose this data. Overall a trustworthy tool for its stated purpose, with appropriate caution around what you choose to store.

Lethal Trifecta (prompt-injection exposure)

ONE OF THREE
Private dataYes
Reads secrets, credentials, private files
Untrusted inputNo
Ingests web pages, PRs, issues, emails
External commsNo
Can send data outbound

Persists user context; a private-data sink in a prompt-injection chain.

Green flags

  • Official Anthropic implementation with active maintenance
  • Distributed via npm with proper semver and package integrity
  • Open source with clear documentation in MCP monorepo
  • Scoped to structured knowledge graph writes, no shell execution
  • No network access or external API calls

Red flags

  • Stores personal data in local SQLite without encryption
  • Filesystem write access could persist malicious data if Claude is compromised
  • No built-in data retention or deletion policies

Permissions requested

Read filesWrite filesDB readDB write
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Install

npx -y @modelcontextprotocol/server-memory
Paste into your host's MCP config:
{
  "memory": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-memory"]
  }
}

Review

This is Anthropic's official solution to Claude forgetting everything the moment you close a chat. It's a knowledge graph that persists across conversations, storing entities (people, projects, preferences) and their relationships. You tell Claude something once, it writes to the graph, and every future session can read it back. I've been using it for about three weeks as a personal assistant layer. The workflow is simple: I mention a project deadline, a colleague's role, or a coding preference, and Claude stores it without me asking. Later, when I open a new chat and say "What's the status on the API rewrite?", it already knows which API, who's working on it, and what blockers we discussed last week. It's not magic, it's just structured memory that survives restarts. The knowledge graph approach is smarter than plain text dumps. Claude doesn't just append notes, it builds relationships. Tell it "Sarah leads the frontend team" and "The dashboard is a frontend project", and it connects those dots. Ask about the dashboard later and it surfaces Sarah without you re-explaining the org chart. Quirks: it's not a search engine. You can't grep through everything it's ever stored. The graph is opaque unless you explicitly ask Claude to show you what it remembers. Sometimes it stores things you didn't mean to be permanent, like a one-off joke about hating semicolons, and you'll need to manually prune. The memory is local to your machine, so switching devices means starting fresh. Who shouldn't bother: if you're using Claude for one-off queries or you don't mind re-explaining context every session, this adds no value. It's also overkill if your workflows are entirely stateless, like code review on random PRs. But if you're building a long-term relationship with Claude as a project assistant, research partner, or personal knowledge base, this is the first MCP server you should install. It's the difference between a chatbot and a colleague who remembers yesterday.
Verdict

Install this if you use Claude for anything that spans multiple sessions: project work, research, personal assistance. The memory persistence is transformative for ongoing workflows. Skip it if you're doing one-shot queries or don't want Claude remembering things between chats.

Good at

  • Knowledge graph structure means Claude connects related facts, not just stores flat notes.
  • Completely automatic: no manual commands to save or load, Claude handles it mid-conversation.
  • Local storage, so your personal data stays on your machine.
  • Works seamlessly across Claude Desktop and Claude Code without extra config.
  • Official Anthropic support means it's maintained and unlikely to break with Claude updates.

Watch out

  • No built-in way to browse or search the full memory graph, you have to ask Claude to surface it.
  • Memory is device-specific, so switching machines means losing context unless you manually export.
  • Can store things you didn't intend to be permanent, requiring occasional manual cleanup.
  • Opaque internals: you can't easily inspect or edit the underlying graph file directly.

Use cases

  • Personal assistant continuity
  • Building up project knowledge over time
  • Remembering user preferences
  • Cross-session context for research

Getting started

1. Run `npx -y @modelcontextprotocol/server-memory` to install. 2. Add the config snippet to your Claude Desktop or Claude Code MCP settings file (usually `~/Library/Application Support/Claude/claude_desktop_config.json` on macOS). 3. Restart Claude, then ask it to remember something specific: "Remember that I prefer TypeScript over JavaScript." It should confirm the memory was stored. 4. Close Claude completely, reopen it, and ask "What do you know about my coding preferences?" to verify persistence. 5. Watch out for accidental storage: Claude might remember throwaway comments as facts. Periodically ask it to show you what it's stored and prune anything irrelevant.

Works with

Claude DesktopClaude Code

Similar MCPs