The Problem
Every new chat starts from scratch. You explain your architecture again. You paste the same patterns again. Your context window bloats with repetition.
Other memory solutions help, but they still require tool calls for every lookup — adding latency and eating into Claude's thinking budget.
Memory MCP fixes this with a two-tier architecture:
- Hot cache (0ms) — Frequently-used knowledge printed into context by the plugin's
SessionStart and UserPromptSubmit hooks before Claude even starts thinking. Claude Code adds a hook's stdout to the conversation, so no tool call is needed.
- Cold storage (~50ms) — Everything else, searchable by meaning via semantic similarity.
The system learns what you use and promotes it automatically. Your most valuable knowledge becomes instantly available. No manual curation required.
Before & After
| 😤 Without Memory MCP | 🎯 With Memory MCP |
|---|
| "Let me explain our architecture again..." | Project facts persist and isolate per repo |
| Copy-paste the same patterns every session | Patterns auto-promoted to instant access |
| 500k+ token context windows | Hot cache keeps it lean (~20 items) |
| Tool call latency on every memory lookup | Hot cache: 0ms — a hook already put it in context |
| Stale information lingers forever | Trust scoring demotes outdated facts |
| Flat list of disconnected facts | Knowledge graph connects related concepts |
Install
uv tool install hot-memory-mcp
claude plugins add michael-denyer/memory-mcp
The plugin gives you auto-configured hooks and slash commands. Embeddings run on sentence-transformers on every platform, including Apple Silicon.
Manual config (no plugin)
Add to ~/.claude.json:
{
"mcpServers": {
"memory": {
"command": "memory-mcp"
}
}
}
See Reference for full configuration options.
Restart Claude Code. Run memory-mcp-cli bootstrap once to seed memories from your project docs; the hooks inject whatever is promoted, but nothing seeds itself.
First run: Embedding model (~90MB) downloads automatically. Takes 30-60 seconds once.
How It Works
flowchart LR
subgraph LLM["Claude"]
REQ((Request))
end
subgraph Hot["HOT CACHE · 0ms"]
HC[Session context]
PM[(Promoted memories)]
end
subgraph Cold["COLD STORAGE · ~50ms"]
VS[(Vector search)]
KG[(Knowledge graph)]
end
REQ -->|"hook stdout"| HC
HC -.->|"draws from"| PM
REQ -->|"recall()"| VS
VS <-->|"related"| KG
The hot cache (~10 items) reaches Claude through two plugin hooks. SessionStart runs memory-mcp-cli hot-cache --force and UserPromptSubmit runs memory-mcp-cli hot-cache, and Claude Code adds each command's stdout to the conversation. The command prints only when the text differs from what the session last saw, which is cheap while the set is stable and reprints in full when it shifts. The set combines recent recalls, predicted next memories, and top promoted items. Promoted memories (~20 items) is the backing store of frequently-used memories. Memories used 3+ times auto-promote; unused ones demote after 14 days.
What Makes It Different
Most memory systems make you pay a tool-call tax on every lookup. Memory MCP's hot cache bypasses this entirely — a hook prints your most-used knowledge into the conversation before Claude starts thinking.
| Memory MCP | Generic Memory Servers |
|---|
| Hot cache | Injected by a hook at 0ms | Every lookup = tool call |
| Self-organizing | Learns and promotes automatically | Manual curation required |
| Project-aware | Auto-isolates by git repo | One big pile of memories |
| Knowledge graph | Multi-hop recall across concepts | Flat list of facts |
| Trust scoring | Outdated info decays and sinks | All memories equal |
| Setup | One command, local SQLite | Often needs cloud setup |
The Engram Insight: Human memory doesn't search — frequently-used patterns are already there. That's what hot cache does for Claude.
Quick Reference
| Slash Command | Tool | Description |
|---|
/memory-mcp:remember | remember | Store a memory with semantic embedding |
/memory-mcp:recall | recall | Search memories by meaning |
/memory-mcp:hot-cache | promote / demote | Manage promoted memories |
/memory-mcp:stats | memory_stats | Show statistics |
/memory-mcp:bootstrap | bootstrap_project | Seed from project docs |
| — | link_memories | Knowledge graph connections |
See Reference for all 10 slash commands and full tool API.
Dashboard
Browse memories, hot cache, injection history, sessions, and the knowledge graph.
How to Use
Memory MCP is designed to run as three complementary components:
| Component | Purpose |
|---|
| Claude Code Plugin | Hooks that inject the hot cache, plus the /memory-mcp:* slash commands |
| MCP Server | Core memory tools available to Claude via Model Context Protocol |
| Dashboard | Web UI to browse, manage, and debug your memory database |
The plugin is recommended for most users — it auto-configures the MCP server and adds productivity features. Run the dashboard alongside when you want visibility into what's being stored.
Documentation
License
MIT
