Memory By File

🚀 mcp-server-memory

This is an MCP server that interacts with a memory text file to help Claude maintain context between chats.

Each line represents a piece of memory.

These tools allow Claude (and other MCP clients) to manage memories during a chat:

• memory_add(memory: string) - Add a memory.
• memory_search(query: string) - Return matching memories (exact substring match). Wildcards/regular expressions may be supported in the future.
• memory_delete(query: string) - Delete matching memories (exact substring match).
• memory_list() - Return all memories.
• Tip: memory_update == memory_delete + memory_add.

For example:

• I mention my name => "Chatting with Wes".
• Mention my daughter's age => "Wes's daughter is 8 years old".
• Say I'm working on a TypeScript project => "Working on a TypeScript project".
• Moreover, it's very important to base on what Claude (the assistant/LLM) says or does...
  • Especially tool usage (i.e., run_command). For example, if the first attempt to use a tool fails (i.e., the python command doesn't exist), and then python3 is used successfully later => Claude can record "Use python3, python doesn't exist".
• I ask Claude to delete memories about X => memory_delete(query: X).
• I correct my name => memory_search("oldname") + delete each matching record or a common subset query + add the new name.

Then, when starting a new chat, Claude will automatically retrieve recent memories (a subset or all), or can request memories (some/more/all). Then, responses/tools, etc., can be influenced based on these memories.

✨ Features

🚀 Quick Start

This mcp-server-memory is designed to assist Claude in maintaining context during chats by interacting with a memory text file. Each line in the file represents a distinct memory.

💻 Usage Examples

Basic Usage

# Add a memory
memory_add("Chatting with Wes")

# Search for memories
matching_memories = memory_search("Wes")

# Delete a memory
memory_delete("old_memory")

# List all memories
all_memories = memory_list()

Advanced Usage

# Update a memory
old_matches = memory_search("oldname")
for match in old_matches:
    memory_delete(match)
memory_add("newname")

🔧 Technical Details

Design

A simple memory text file is used for the following reasons:

• ChatGPT's memory works well and is essentially a text file.
  • There may be some structured processing behind it, but when you view the memory, it appears as a text file.
• It worked well in my tests of the mcp-server-commands with a similar reminder function (Claude had these functions at that time).
• Unstructured text simplifies tool operations and parameters, basically just managing a list of strings.

Prompt Mechanism

• It's also important to have a prompt to let the model know when to store memories. It's not clear yet how to best implement this...
• Training: OpenAI acknowledges that some models during training know when to store memories. Just like models are trained for tool usage.
• Prompt words: A system prompt section may contain a reminder to trigger memory storage.
• The tool itself: In my tests with Claude, when only using prompts/suggestions in the tool specification and tool responses, Claude still couldn't store memories. So, relying solely on the tool is not enough. I think adding a prompt in the prompt will be effective.

📄 Unfinished Ideas

I'm not sure if these are worth the time. I'm just listing some ideas for future implementation. Partly to prevent myself from getting distracted by implementing them :)

• Novelty factor: A way to rearrange memories based on novelty?
  • Chronological order becomes important for ambiguous memory queries (e.g., if both a TypeScript project and a Python project are mentioned, and I ask to start a new project, it may suggest the one used most recently).
• Gradually fade out old memories?