# AgenticMemory MCP Server
> Universal LLM access to persistent graph memory via the Model Context Protocol.
AgenticMemory MCP Server bridges **any MCP-compatible LLM client** (Claude, GPT, Gemini, Ollama, etc.) to the AgenticMemory persistent binary graph memory system. One server, universal access.
## Architecture
```
┌─────────────────────────────────────────────────────────┐
│ LLM CLIENT │
│ (Claude, GPT, Gemini, Ollama, etc.) │
└─────────────────────┬───────────────────────────────────┘
│ MCP Protocol (JSON-RPC 2.0)
│ (stdio / SSE)
┌─────────────────────▼───────────────────────────────────┐
│ AGENTIC-MEMORY-MCP │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ TOOLS │ │RESOURCES │ │ PROMPTS │ │
│ │ 12 tools │ │ 6 URIs │ │ 4 tmpls │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ └──────────────┼─────────────┘ │
│ SESSION MANAGER │
│ AGENTIC-MEMORY CORE │
└─────────────────────────────────────────────────────────┘
│
▼
brain.amem (binary graph file)
```
## Quick Start
### Install from crates.io
```bash
cargo install agentic-memory-mcp
```
### Or build from source
```bash
cargo build --release
```
### Run (stdio transport — default)
```bash
agentic-memory-mcp serve
```
> Zero-config: defaults to `~/.brain.amem`. Override with `agentic-memory-mcp --memory /path/to/brain.amem serve`.
### Memory Modes
Control how aggressively Claude saves memories:
```bash
agentic-memory-mcp serve --mode minimal # Only save when user says "remember"
agentic-memory-mcp serve --mode smart # Auto-save facts + decisions (default)
agentic-memory-mcp serve --mode full # Save everything, episode summaries
```
| **minimal** | 1-3 | ~1 KB/day | Privacy-conscious users |
| **smart** | 5-15 | ~5-20 KB/day | Most users (default) |
| **full** | 15-30+ | ~20-50 KB/day | Power users, autonomous agents |
Configure via the `args` array in your MCP config:
```json
{ "command": "agentic-memory-mcp", "args": ["serve", "--mode", "full"] }
```
### Run (SSE transport)
```bash
cargo build --release --features sse
agentic-memory-mcp serve-http --addr 127.0.0.1:3000
```
## Configuration with MCP Clients
### Claude Desktop
Add to `~/Library/Application Support/Claude/claude_desktop_config.json`:
```json
{
"mcpServers": {
"agentic-memory": {
"command": "agentic-memory-mcp",
"args": ["serve"]
}
}
}
```
### Claude Code
Add to `~/.claude/mcp.json`:
```json
{
"mcpServers": {
"agentic-memory": {
"command": "agentic-memory-mcp",
"args": ["serve"]
}
}
}
```
### VS Code / Cursor
Add to `.vscode/settings.json`:
```json
{
"mcp.servers": {
"agentic-memory": {
"command": "agentic-memory-mcp",
"args": ["serve"]
}
}
}
```
### Windsurf
Add to `~/.codeium/windsurf/mcp_config.json`:
```json
{
"mcpServers": {
"agentic-memory": {
"command": "agentic-memory-mcp",
"args": ["serve"]
}
}
}
```
> Zero-config: defaults to `~/.brain.amem`. Override with `"args": ["--memory", "/path/to/brain.amem", "serve"]`.
## Tools (12)
| `memory_add` | Add cognitive event to graph |
| `memory_query` | Pattern query for matching nodes |
| `memory_traverse` | Walk graph following edges |
| `memory_correct` | Record correction to past belief |
| `memory_resolve` | Follow supersedes chain |
| `memory_context` | Get subgraph around node |
| `memory_similar` | Similarity search |
| `memory_causal` | Impact analysis |
| `memory_temporal` | Compare across time |
| `memory_stats` | Graph statistics |
| `session_start` | Begin new session |
| `session_end` | End session, create episode |
## Resources (6)
| `amem://node/{id}` | Single node with edges |
| `amem://session/{id}` | All nodes from session |
| `amem://types/{type}` | All nodes of type |
| `amem://graph/stats` | Graph statistics |
| `amem://graph/recent` | Recent nodes |
| `amem://graph/important` | High decay score nodes |
## Prompts (4)
| `remember` | Guide for storing new info | `information` |
| `reflect` | Guide for understanding past decisions | `topic` |
| `correct` | Guide for updating beliefs | `old_belief`, `new_information` |
| `summarize` | Guide for session summary | (none) |
## Event Types
- **Fact** — Declarative knowledge
- **Decision** — Choices made with reasoning
- **Inference** — Derived conclusions
- **Correction** — Updates to prior beliefs
- **Skill** — Procedural knowledge
- **Episode** — Session summaries
## Edge Types
- **CausedBy** — Causal relationship
- **Supports** — Supporting evidence
- **Contradicts** — Conflicting information
- **Supersedes** — Correction chain
- **RelatedTo** — General association
- **PartOf** — Hierarchical containment
- **TemporalNext** — Temporal ordering
## CLI Commands
```bash
# Start server (stdio) — defaults to ~/.brain.amem
agentic-memory-mcp serve
# Start server with custom memory file
agentic-memory-mcp --memory /path/to/brain.amem serve
# Start server (SSE, requires --features sse)
agentic-memory-mcp serve-http --addr 127.0.0.1:3000
# Validate a memory file
agentic-memory-mcp --memory ~/.brain.amem validate
# Print server info as JSON
agentic-memory-mcp info
```
## Development
This crate is part of the [AgenticMemory](../../README.md) Cargo workspace.
```bash
# Run MCP server tests (from workspace root)
cargo test -p agentic-memory-mcp
# Run bridge integration tests
cargo test -p agentic-memory-bridge-tests
# Run all workspace tests
cargo test --workspace
# Clippy + format
cargo clippy --workspace
cargo fmt --all
# Build release
cargo build --release
```
## Protocol
This server implements MCP (Model Context Protocol) spec version **2024-11-05** over JSON-RPC 2.0. Supported transports:
- **stdio** — Newline-delimited JSON over stdin/stdout (default)
- **SSE** — Server-Sent Events over HTTP (feature flag `sse`)
## License
MIT