Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@mcp-memory recall what we discussed about the deployment"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Your AI agent forgets everything between sessions. This fixes that.
An MCP server that gives any AI agent persistent memory with semantic search. Store decisions, context, and knowledge once -- recall them with natural language queries across any future session.
Built on ChromaDB embeddings, scoped per project, runs entirely locally.
Every MCP-based agent (Claude Desktop, Claude Code, Cursor) starts each session with amnesia. Decisions made yesterday are gone. Context from last week is gone. You re-explain the same things every time.
mcp-memory adds four tools -- remember, recall, forget, list_memories -- that persist knowledge across sessions with vector similarity search. Your agent remembers what matters and finds it when relevant.
Related MCP server: Synapto
Semantic recall -- vector embeddings (all-MiniLM-L6-v2) find related memories, not just keyword matches
Per-project scoping -- memories don't leak between projects
Importance scoring -- prioritize critical decisions (1-5 scale)
Tag-based filtering -- organize memories by category
Fully local -- ChromaDB on disk, no cloud, no API keys, no telemetry
pip install -e .Environment Variable | Default | Description |
|
| Where memories are stored on disk |
|
| Default project scope |
|
| Default number of recall results |
Add to ~/.config/claude/claude_desktop_config.json:
{
"mcpServers": {
"memory": {
"command": "mcp-memory",
"env": {
"MCP_MEMORY_DATA_DIR": "~/.mcp-memory"
}
}
}
}Add to .claude/settings.json:
{
"mcpServers": {
"memory": {
"command": "mcp-memory"
}
}
}Store a memory for later recall.
Arg | Type | Default | Description |
| string | required | The text to remember |
| string | "global" | Project scope |
| list[string] | [] | Tags for filtering |
| string | "" | Where this memory came from |
| int | 3 | Priority 1-5 |
Search memories by semantic similarity.
Arg | Type | Default | Description |
| string | required | Natural language search |
| string | all | Limit to project |
| list[string] | none | Filter by tags |
| int | 10 | Max results |
| float | none | Minimum relevance 0.0-1.0 |
Delete stored memories.
Arg | Type | Default | Description |
| list[string] | none | Specific IDs to delete |
| string | none | Delete all in project |
| list[string] | none | Delete by tags |
Browse stored memories with pagination.
Arg | Type | Default | Description |
| string | all | Filter to project |
| list[string] | none | Filter by tags |
| int | 1 | Page number |
| int | 20 | Results per page |
pip install -e ".[dev]"
pytest # run tests
ruff check . # lint
ruff format . # format
mypy mcp_memory # type checkMIT
Delete stored memories. Specify at least one filter criterion.
Inputs: tags, project, memory_ids. Returns: result
Browse and list stored memories with optional filtering.
Inputs: page, tags, project, page_size. Returns: result
Search memories by semantic similarity.
Returns the most relevant stored memories.
Inputs: query (required), tags, project, n_results, +1 more. Returns: result
Store a memory for later recall. Content is embedded for semantic search.
Inputs: content (required), tags, source, project, +1 more. Returns: result
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 4 tools
Each tool has a distinct purpose: remember stores, recall searches semantically, list_memories browses with filtering, and forget deletes. There is no overlap or ambiguity.
Most tools use single verb names (remember, recall, forget), but 'list_memories' breaks the pattern with a verb_noun style. The naming is clear but not fully consistent.
With 4 tools covering create, read (two retrieval methods), and delete, the count is well-scoped for a memory server. No tools feel redundant or missing.
The set covers create, read (list and search), and delete, but lacks an update operation. There is no way to modify an existing memory, which is a notable gap for a typical CRUD lifecycle.
Persistent memory for AI agents across Claude, ChatGPT and any MCP client.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Persistent memory for AI agents to retain, retrieve, and recall conversation context through MCP.
Hosted memory for AI agents over MCP. One memory shared by all your agents.
