Claude Memory Tools

A forked and improved version of @whenmoon-afk/memory-mcp with better FTS5 search, plus Python utilities for importing data into the memory database.

What's different from upstream

The upstream MCP passes search queries directly to SQLite FTS5 with no preprocessing. FTS5's default operator is AND, so a query like "bonsai tree plant care" requires all four words in a single memory — returning zero results even when relevant memories exist.

This fork adds:

• OR expansion — multi-word queries are split and joined with OR, so any matching term returns results
• Prefix wildcards — UMass matches UMass Global, home matches homelab
• Quoted phrase passthrough — "exact phrase" still does exact matching
• Term-match boosting — memories matching more of your query terms rank higher

Query	Upstream	This fork
bonsai tree plant care	0 results	89 results
plant care	2 results	84 results
book writing style	22 results	287 results

The existing porter stemming and hybrid scoring (importance/recency/frequency) are preserved.

Setup

MCP Server (the fork)

Install dependencies:

cd memory-mcp-fork
npm install --ignore-scripts

Claude Code (~/.claude/.mcp.json)

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["/path/to/claude-memory-tools/memory-mcp-fork/dist/index.js"]
    }
  }
}

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["/path/to/claude-memory-tools/memory-mcp-fork/dist/index.js"]
    }
  }
}

Python Tools

No dependencies beyond Python 3.10+ standard library.

Tools

import_to_memory.py — Import markdown files

Import a folder of .md files (e.g. Capacities/Obsidian/Notion exports) as memories:

python import_to_memory.py /path/to/markdown/folder

• Each file becomes one memory entry
• Generates summaries from content
• Populates FTS index
• Stores source filename in metadata for deduplication

import_all_sessions.py — Import Claude Code sessions

Batch import all Claude Code JSONL session files as searchable memories:

python import_all_sessions.py                  # convert + import all
python import_all_sessions.py --rebuild-fts    # rebuild FTS index only
python import_all_sessions.py --convert-only   # dry run, no DB changes
python import_all_sessions.py --import-only    # import pre-converted chunks

• Chunks large sessions into ~4000 char pieces (fits MCP's token budget)
• Generates human-readable summaries
• Deduplicates by source filename
• Skips agent subprocesses and tiny sessions (< 5 messages)

memory_search.py — CLI search

Search, list, and manage memories from the command line:

python memory_search.py bonsai              # search
python memory_search.py "plant care"        # phrase search
python memory_search.py --list              # list all
python memory_search.py --stats             # DB statistics
python memory_search.py --delete mem_abc    # soft-delete by ID prefix

export_chat.py — Export conversations

Convert a Claude Code JSONL session to readable markdown:

python export_chat.py session.jsonl                    # outputs session.md
python export_chat.py session.jsonl my_export.md       # custom output path

Database location

The MCP stores its SQLite database at:

Platform	Path
Windows	%APPDATA%/claude-memories/memory.db
macOS	~/.claude-memories/memory.db
Linux	~/.local/share/claude-memories/memory.db

Override with the MEMORY_DB_PATH environment variable.

License

The memory-mcp-fork is based on @whenmoon-afk/memory-mcp (MIT license). Python tools are MIT licensed.