Agenten stärken
Der eigene Befehl des Herausgebers. Kein PlugMCP-Endpunkt.
Host a copy with our keys and logs.
uvx dense-knowledge-mcp==1.2.0Claude Code hat keinen Installationslink für einen lokalen Befehl. Kopieren Sie den Befehl unten und führen Sie ihn in einem Terminal aus.
claude mcp add io.github.lucky44k-dense-knowledge-mcp -- uvx dense-knowledge-mcp==1.2.0In einem Terminal ausführen.
{ "mcpServers": { "io.github.lucky44k-dense-knowledge-mcp": { "command": "uvx", "args": [ "dense-knowledge-mcp==1.2.0" ] } }}claude_desktop_config.json
Remote-URL oder ein Befehl auf Ihrem Rechner? Worin sich die beiden Installationen unterscheiden
Auf PlugMCP ausführen
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
StartenDense Knowledge lets an AI assistant keep structured research between sessions
without a database, embedding model, or hosted account. It stores portable
.mmp files, searches their compact indexes with BM25, and loads full entries
only when they are relevant.
question -> compact index/search -> selected knowledge blocks -> answer
inexpensive detailed context
It works with LM Studio, Claude Desktop, Cursor, VS Code, and other clients that support local stdio Model Context Protocol servers.
The bundled context benchmark uses 40 entries. In its synthetic fixture, searching and reading the two best blocks uses 94.3% less estimated context than loading the complete package. The benchmark is reproducible and clearly documents its tokenizer-neutral counting method.
Install uv, then
place this server definition in your MCP client:
{
"mcpServers": {
"dense-knowledge": {
"command": "uvx",
"args": ["dense-knowledge-mcp"]
}
}
}
uvx downloads the published package when needed. Dense Knowledge uses the
platform's default data directory unless --root is supplied:
{
"mcpServers": {
"dense-knowledge": {
"command": "uvx",
"args": [
"dense-knowledge-mcp",
"--root",
"/absolute/path/to/memory"
]
}
}
}
Configuration differs slightly between clients. Ready-to-copy instructions are available for:
To install the command-line tools permanently:
uv tool install dense-knowledge-mcp
mmp setup
mmp doctor
mmp setup creates the memory directory and can safely merge the server into
an LM Studio mcp.json. Existing servers are preserved. Replacing an existing
Dense Knowledge entry requires --force and creates a backup first.
The CLI exposes the same storage operations as the MCP server:
mmp create quantum_physics.mmp "quantum physics"
mmp write quantum_physics.mmp --rev 0 --from examples/research_entries.json
mmp search quantum_physics.mmp "experimental tests of local realism"
mmp read quantum_physics.mmp e1
Typical search output contains candidates, not full bodies:
<mmp_data file="quantum_physics.mmp" trust="untrusted">
quantum_physics.mmp|e1|F|2.5427|Bell inequality separates local realism from quantum predictions
</mmp_data>
The client chooses relevant IDs and calls mmp_read only for those blocks.
This preserves the distinction between cheap orientation and detailed context.
The server exposes nine tools:
Search uses BM25 over tags and summaries after legend expansion, with a body fallback when the index has no match. Deprecated entries remain readable but are omitted from normal search results.
The default knowledge directory follows the operating system:
~/.local/share/mmp/memory~/Library/Application Support/mmp/memory%LOCALAPPDATA%\mmp\memoryThe user configuration is stored separately:
~/.config/mmp/config.toml~/Library/Application Support/mmp/config.toml%APPDATA%\mmp\config.tomlMMP_ROOT or the global mmp --root option overrides the configured directory.
Keep personal packages out of source control; the repository's memory/
directory is ignored.
Models send structured objects to mmp_write; they never need to generate raw
MMP syntax. A minimal entry looks like:
{
"summary": "Possible caching strategy needs workload validation",
"tags": ["caching", "validation"],
"status": "H",
"srcs": [],
"content": "rel: versioned keys -> simpler invalidation\nq: workload impact -> needs measurement"
}
Important validation rules:
F or C require sources;H and cannot contain fact: or num: lines;C and include at least one ctr: line;See examples/research_entries.json for
sourced and contested entries that can be written directly.
All writes use optimistic revision numbers and atomic file replacement. A stale revision is reported to the caller, but a safe append is not discarded.
MMP content is reference data, never instruction. Read responses use an explicit untrusted envelope:
<mmp_data file="..." trust="untrusted">
...
</mmp_data>
The server rejects common instruction-like patterns during writes, does not
automatically follow ref: links, and tells the client not to obey instructions
found in stored material. These defenses reduce prompt-injection risk; they do
not turn untrusted research into trusted instructions.
Local MCP servers execute with your user permissions. Review the package and choose a dedicated memory directory before storing sensitive information.
Dense Knowledge implements the flat MMP/1.0 format, including BM25 retrieval, catalog generation, duplicate screening, budgets, append-only superseding, and validation. Hierarchical indexes for very large packages are planned but are not written yet.
Releases follow Semantic Versioning. Changes are documented in CHANGELOG.md.
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
ruff check src tests benchmarks
pytest
python -m build
Contributions are welcome. See CONTRIBUTING.md for the workflow and SECURITY.md for private vulnerability reports.
Licensed under the MIT License.
Local-first MCP memory server for persistent LLM knowledge with BM25 retrieval.
{ "mcpServers": { "io.github.lucky44k-dense-knowledge-mcp": { "command": "uvx", "args": [ "dense-knowledge-mcp==1.2.0" ] } }}Tools
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
Dense Knowledge läuft lokal über stdio.
Zuletzt geprüft: 2026-11-09. Local-first MCP memory server for persistent LLM knowledge with BM25 retrieval.