Mem0

Hosted long-term memory for your agent — searchable facts that persist across sessions, extracted automatically from conversation.

What it installs

Artifact	Path (in your project)	Purpose
MCP server	`.mcp.json` → `mem0` entry	Connects the agent to mem0 cloud — 8+ memory tools (`search_memories`, `add_memory`, `update_memory`, …)
Skill	`.claude/skills/mem0/`	Protocol the agent follows: search-before-answer, scope-by-user-id, let mem0 extract facts from raw messages

How it works

mem0 is a managed memory layer. You send it the messages[] of a turn; its LLM extracts the salient facts, dedupes them against existing memories, and indexes for semantic + graph retrieval. On the next turn your agent calls search_memories(query) to pull the top-K relevant facts and injects them into reasoning.

The skill teaches the agent the workflow that separates useful memory from noise:

Always scope by user_id — without it, every user's memories land in the same bucket.
Search every turn, not just session start — memories mutate.
Pass raw messages to add_memory, never pre-summarized text — mem0's extractor is the whole point.
Add at end-of-turn — one extraction call per resolved task.

Without the skill, agents commonly: forget user_id, store raw transcripts (bloat), or only search on session start (stale recall).

Setup

Sign up at https://app.mem0.ai and create an API key at https://app.mem0.ai/settings/api-keys.
Export in your shell or .env:
```
MEM0_API_KEY=m0-...
```
Restart Claude Code so the MCP server picks up the env var.

No Docker, no self-hosted infra — this bundle targets the hosted platform. Self-hosted mem0 (BYO LLM + vector store) is possible but not covered here.

Pairs well with

local-memory — mem0 is cloud-hosted and team-shareable via API scope; local-memory is file-based and repo-committable. Pick one primary memory system; using both fragments recall.
context-discipline — mem0 reduces the need to keep facts hot in context because retrieval is cheap. The two together let you run leaner system prompts.