Skip to content
MCP server

Polypack Memory

Persistent graph-based memory for MCP clients, stored locally and shared across agents

Runs locally
Install

Polypack Memory is an MCP server for knowledge and memory. It runs locally on your machine and is started by your AI client.

Platforms
CLI, Self-hosted, Linux
Runs
Locally
Version
0.1.32
Listing updated

Similar MCP servers

Install

Run in a terminal

claude mcp add polypack-mcp -- uvx polypack-mcp

Add to claude_desktop_config.json

{
  "mcpServers": {
    "polypack-mcp": {
      "command": "uvx",
      "args": [
        "polypack-mcp"
      ]
    }
  }
}
Install in CursorOpens Cursor, which asks you to confirm. Or add it by hand:

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "polypack-mcp": {
      "command": "uvx",
      "args": [
        "polypack-mcp"
      ]
    }
  }
}
Install in VS CodeOpens VS Code, which asks you to confirm. Or add it by hand:

Add to .vscode/mcp.json

{
  "servers": {
    "polypack-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "polypack-mcp"
      ]
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "polypack-mcp": {
      "command": "uvx",
      "args": [
        "polypack-mcp"
      ]
    }
  }
}

Add to your client's MCP settings (most use this shape)

{
  "mcpServers": {
    "polypack-mcp": {
      "command": "uvx",
      "args": [
        "polypack-mcp"
      ]
    }
  }
}

Using several servers? Build one config file with the MCP config generator.

About Polypack Memory

Polypack Memory keeps what an assistant learns in a local Polypack store so Claude, Codex and other clients can share durable memory. For one client it runs over stdio; for several it runs as a long-lived local HTTP service at 127.0.0.1:8765, set up with polypack-mcp setup as a systemd user service or installed from Debian and RPM packages. The README warns not to start two processes on the same store.

Seventeen tools cover storing, updating, linking, recalling and pruning memories. Memories fall into entity, episodic, procedural and semantic classes, with preferences stored as procedural. Recall mixes lexical and activation scores and can add optional local semantic retrieval using Qwen3-Embedding-0.6B, which is downloaded once and unloaded after 15 minutes idle. Retrieval can pull in neighbouring memories along graph edges, and a token budget caps what is returned. Feedback reinforces useful memories, while supersession and consolidation keep history through graph edges.

Context works as a soft preference, with strict_context available for hard isolation. Deleting is permanent and needs confirmation, and memory_suppress retains history instead.

Tools this server provides (17)

Taken from the project documentation; the running server may differ.

memory_store
Stores a memory.
memory_get
Fetches a memory by exact id.
memory_update
Updates context, confidence, provenance or metadata.
memory_list_contexts
Lists context namespaces.
memory_delete
Permanently deletes a memory after confirmation.
memory_recall
Recalls memories by query with optional neighbours.
memory_context
Returns memories within a token budget.
memory_feedback
Records whether a memory was useful.
memory_suppress
Suppresses a memory while keeping history.
memory_supersede
Replaces a memory's content while keeping history.
memory_consolidate
Consolidates memories.
memory_link
Links two memories.
memory_unlink
Removes a link between memories.
memory_thread
Reads a thread of linked memories.
memory_store_batch
Stores several memories at once.
memory_link_batch
Links several memories at once.
graph_query
Runs graph queries such as relationship diagnostics.

What it does

Durable local store
Mutations checkpoint immediately and the store flushes on shutdown.
Shared HTTP mode
One local service lets Claude and Codex use the same memory.
Graph links
Link, unlink and traverse memories, with neighbours returned on recall.
Adaptive recall
Scores combine lexical, activation and optional semantic matches.
Token-budgeted context
memory_context never returns an item that would exceed the budget.
History-preserving edits
Supersede and consolidate memories without losing earlier versions.
Context scoping
Prefer or strictly isolate memories by context namespace.

Good for

  • Give Claude and Codex a common long-term memory on one machine
  • Store project conventions as procedural memories
  • Record a fix and link it to the earlier memory it answers
  • Turn on local semantic search without a hosted embedding service
  • Retire an outdated fact while keeping its history

Polypack Memory FAQ

How do I install Polypack Memory?

pip install 'polypack-mcp[polypack]', or use the Debian, RPM or APT repository packages.

Can two clients share memory?

Yes, run the shared Streamable HTTP service and give both clients its URL instead of a command.

Is a model required?

No. The default uses graph, activation and lexical retrieval; Qwen embeddings are optional.

How many tools are there?

Seventeen, from memory_store to graph_query.

Is deletion reversible?

No. memory_delete is permanent and needs confirm=true; memory_suppress keeps history.

How do I add Polypack Memory to Claude Code?

Run this command in your terminal: claude mcp add polypack-mcp -- uvx polypack-mcp

What are the best alternatives to Polypack Memory?

The closest alternatives to Polypack Memory are Compound Memory, VelesDB Memory, persistent-memory, Memory MCP and Wontopos.

Where this comes from

  • Description: written from the project's README.
  • Install steps: from the official MCP Registry entry.

How AIPoweredBot checks servers