About Memento
Memento is a local MCP memory server for users who need persistent context across conversations. It stores a structured knowledge graph, supports offline BGE-M3 semantic search, and offers SQLite or PostgreSQL storage.
Key features
- 01Entity, observation, and relation graph
- 02Offline BGE-M3 semantic search
- 03SQLite and PostgreSQL backends
- 04Multi-factor relevance scoring
- 05Memory importance settings
- 06JSONL write-event hook
Use cases
- 01Persisting and retrieving context across conversations
- 02Finding memory nodes by meaning
- 03Forwarding memory changes to other processes
MCP support
The MCP server exposes tools to create, search, read, and delete graph data, manage observations and importance, and optionally emit JSONL events after mutations.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEMemento
Some memories are best persisted.
Provides persistent memory capabilities through a SQLite-based knowledge graph that stores entities, observations, and relationships with semantic search using BGE-M3 embeddings for intelligent context retrieval across conversations.
Features
- Semantic vector search (sqlite-vec/pgvector, 1024d)
- Offline embedding model (
bge-m3) - Modular repository layer with SQLite and PostgreSQL backends
- Enhanced Relevance Scoring with temporal, popularity, contextual, and importance factors
- Structured graph of
entities,observations, andrelations - Easy integration with Claude Desktop (via MCP)
Prerequisites
System SQLite Version Check
Memento requires SQLite 3.38+. Most macOS and Linux distros ship sqlite3 out of the box, but double-check that it's there and new enough:
sqlite3 --version # should print a version string, e.g. 3.46.0
Important Note: This check is just to verify SQLite is installed on your system. Memento does NOT use the sqlite3 CLI for its operation it uses the Node.js sqlite3 module internally.
If you see "command not found" (or your version is older than 3.38), install SQLite:
| Platform | Install command |
|---|---|
| macOS (Homebrew) | brew install sqlite |
| Debian / Ubuntu | sudo apt update && sudo apt install sqlite3 |
Configuration
Memento now supports pluggable storage backends. Configuration is controlled entirely through environment variables so it remains easy to embed inside MCP workflows.
| Variable | Description |
|---|---|
MEMORY_DB_DRIVER | Optional selector for the database backend. Defaults to sqlite. Set to postgres to enable the PostgreSQL manager. |
MEMORY_DB_PATH | Filesystem path for the SQLite database file (only used when the driver is sqlite). |
SQLITE_VEC_PATH | Optional absolute path to a pre-built sqlite-vec extension shared library. |
MEMORY_DB_DSN / DATABASE_URL | PostgreSQL connection string consumed by the pg client. |
PGHOST, PGPORT, PGUSER, PGPASSWORD, PGDATABASE | Individual PostgreSQL connection parameters. Used when no DSN is provided. |
PGSSLMODE | When set to require, SSL will be enabled with rejectUnauthorized: false. |
MEMENTO_WRITE_HOOK_PATH | Optional path to a JSON Lines file. When set, Memento appends one event after each successful graph mutation. |
Write hook events
Set MEMENTO_WRITE_HOOK_PATH when another process needs to react to memory writes without polling the full graph:
MEMENTO_WRITE_HOOK_PATH=/tmp/memento-events.jsonl memento
Each line is a JSON object with an ISO timestamp, an operation, and the changed entity/relation payload. Example:
{"timestamp":"2026-05-05T16:33:16.619Z","operation":"create_entity","entity":{"name":"Hook Test","entityType":"test"}}
Supported operations are create_entity, add_observations, create_relation, delete_entities, delete_relations, delete_observations, and set_importance. Delete events are emitted only for records that were actually removed.
Hook write failures are logged to stderr but do not fail the graph mutation that already succeeded. Consumers that require strict reconciliation should combine the hook with occasional read_graph or targeted open_nodes checks.
PostgreSQL notes
- The PostgreSQL manager requires the
pgvectorextension. It is automatically initialized withCREATE EXTENSION IF NOT EXISTS vector.
Claude Desktop:
{
"mcpServers": {
"memory": {
"description": "Custom memory backed by SQLite + vec + FTS5",
"command": "npx",
"args": [
"@iachilles/memento@latest"
],
"env": {
"MEMORY_DB_PATH": "/Path/To/Your/memory.db"
},
"options": {
"autoStart": true,
"restartOnCrash": true
}
}
}
}
Troubleshooting
sqlite-vec Extension Issues
Important: Memento loads the sqlite-vec extension programmatically through Node.js, NOT through the sqlite3 CLI.
Common misconceptions:
- ❌ Creating shell aliases for sqlite3 CLI won't affect Memento
- ❌ Loading extensions in sqlite3 CLI won't help Memento
- ✅ Use the npm-installed sqlite-vec or set
SQLITE_VEC_PATHenvironment variable if automatic detection fails. This should point to the Node.js-compatible version of the extension, typically found in yournode_modulesdirectory.
If automatic vec loading fails:
# Find the Node.js-compatible vec extension
find node_modules -name "vec0.dylib" # macOS
find node_modules -name "vec0.so" # Linux
# Use it via environment variable
SQLITE_VEC_PATH="/full/path/to/node_modules/sqlite-vec-darwin-x64/vec0.dylib" memento
API Overview
This server exposes the following MCP tools:
create_entitiescreate_relationsadd_observationsdelete_entitiesdelete_relationsdelete_observationsread_graphsearch_nodesopen_nodesset_importance- Set importance level (critical/important/normal/temporary/deprecated)
When MEMENTO_WRITE_HOOK_PATH is configured, mutating tools also append JSONL write events for downstream consumers.
An example of an instruction set that an LLM should know for effective memory handling (see MEMORY_PROTOCOL.md)
Embedding Model
This project uses @xenova/transformers, with a quantized version of bge-m3, running fully offline in Node.js.
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.