memory_bootstrapLABSCO SUMMARY
Memory you can open in a text editor.
The design choice that matters here is storage: memory is Markdown on disk, so you can read it, edit it, and check it into a repo, and the SQLite index is rebuilt around it rather than being the source of truth. Because the workspace is a directory rather than a hosted account, several clients can share one identity, and a second project just means a second directory.
What it is
A local-first memory server that gives an agent long-term facts, a user profile, an entity graph and daily logs that survive between sessions. Memory lives in readable Markdown — MEMORY.md, USER.md, AGENTS.md, RELATIONS.md and dated files under daily/ — indexed in SQLite so the same workspace can be attached to Claude Desktop, Cursor, Cline or Open WebUI at once.
What you get
- Load the whole memory context at the start of a session — long-term facts, user profile, agent instructions, relations and daily logs — with
memory_bootstrap - Write a fact to a chosen tier with
memory_write, so preferences and decisions go to long-term memory and progress notes go to the daily log - Search across tiers together or one at a time with
memory_read, combining BM25 keyword scoring and vector similarity in a single ranked query - Record typed relationships between entities with
memory_relate, with near-duplicate triples suppressed by cosine similarity - An optional
memory_listtool, and a dispatcher mode that routes every action through a singlememory_toolcall for clients that do better with a small tool surface
Requirements
Nothing by default: with provider: none it runs on SQLite with FTS5 alone — no API key, no GPU, no extra dependencies. Python 3.12 or newer; the package is groundmemory (version 0.6.13) and groundmemory-mcp serves HTTP at http://127.0.0.1:4242/mcp, with stdio also supported. docker compose up -d gives the same endpoint. Semantic search is optional and takes a local sentence-transformers model or any OpenAI-compatible endpoint, configured through GROUNDMEMORY_EMBEDDING__PROVIDER and friends.
Setup effort
One command — pip install groundmemory && groundmemory-mcp
USE CASES
01
Carrying context between two different editors
Point Claude Desktop and Cursor at the same workspace, then have each session reload the stored profile and facts instead of asking again.
02
Keeping decisions from being re-litigated
Write architectural decisions and stated preferences into long-term memory so later sessions retrieve them instead of proposing alternatives you already rejected.
03
Building a picture of who relates to what
Record typed triples as a research agent reads sources, then query the graph later without rereading the material.
04
Running memory with no cloud dependency
Keep the BM25-only mode on an offline machine, so recall works with no embedding provider and no network call.
5 tools
Listed in project docs. Your client may show a different list.
memory_readmemory_writememory_relatememory_listBEFORE YOU INSTALL
What to know first
Tools you get
5 tools named in the project's documentation.
Where it runs
All 4 documented clients run the server on your machine.
INSTALLATION
Connect your client
Server setup
TERMINAL
pip install groundmemory && groundmemory-mcpMCP Clients
Cursor, Other MCP clients
JSON
{
"mcpServers": {
"GroundMemory": {
"url": "http://127.0.0.1:4242/mcp"
}
}
}Cline
CLINE · JSON
{
"mcpServers": {
"GroundMemory": {
"type": "streamableHttp",
"url": "http://127.0.0.1:4242/mcp"
}
}
}Windsurf
WINDSURF · JSON
{
"mcpServers": {
"GroundMemory": {
"serverUrl": "http://127.0.0.1:4242/mcp"
}
}
}
