Description:
Hi MemPalace team! First of all, thanks for the amazing work on this project. The retrieval accuracy in raw mode is incredibly impressive.
I'm currently using MemPalace with Claude Code, and I wanted to discuss the user experience regarding the mempal-stop-hook.sh auto-save mechanism.
Current Behavior:
When exiting a Claude Code session, the stop hook blocks the exit ("decision": "block") and forcefully injects the following prompt into the terminal:
AUTO-SAVE checkpoint. Save key topics, decisions, quotes...
Because of how Claude Code handles standard output/error from hooks, this often surfaces with a jarring Stop hook error: prefix.
The Problem:
I understand why it's built this way: the hook needs to inject this prompt into the current context so the LLM actually performs the summarization and triggers the MCP tool before dying. However, it is very noisy and breaks the developer's flow. If a user tries to manually silence the shell script (e.g., using sed to replace "block" with "continue"), the prompt is never processed by the LLM, and the auto-save functionality breaks entirely.
Feature Request:
It would be great to have a --silent or --quiet configuration for the hooks.
Proposed Implementation Ideas:
Background Processing (Ideal): Instead of forcing the active terminal LLM to read a prompt, the hook could capture the current session log and pass it to a background Python daemon. That daemon could then make its own silent LLM API call to summarize and ingest the data into ChromaDB without blocking the user's UI.
Clean Opt-out: Provide an official configuration toggle in mempalace config to completely disable the auto-save hook for users who prefer to trigger memory storage manually (e.g., using mempalace mine --mode convos later).
Thanks for considering this!
Description:
Hi MemPalace team! First of all, thanks for the amazing work on this project. The retrieval accuracy in raw mode is incredibly impressive.
I'm currently using MemPalace with Claude Code, and I wanted to discuss the user experience regarding the mempal-stop-hook.sh auto-save mechanism.
Current Behavior:
When exiting a Claude Code session, the stop hook blocks the exit ("decision": "block") and forcefully injects the following prompt into the terminal:
AUTO-SAVE checkpoint. Save key topics, decisions, quotes...
Because of how Claude Code handles standard output/error from hooks, this often surfaces with a jarring Stop hook error: prefix.
The Problem:
I understand why it's built this way: the hook needs to inject this prompt into the current context so the LLM actually performs the summarization and triggers the MCP tool before dying. However, it is very noisy and breaks the developer's flow. If a user tries to manually silence the shell script (e.g., using sed to replace "block" with "continue"), the prompt is never processed by the LLM, and the auto-save functionality breaks entirely.
Feature Request:
It would be great to have a --silent or --quiet configuration for the hooks.
Proposed Implementation Ideas:
Background Processing (Ideal): Instead of forcing the active terminal LLM to read a prompt, the hook could capture the current session log and pass it to a background Python daemon. That daemon could then make its own silent LLM API call to summarize and ingest the data into ChromaDB without blocking the user's UI.
Clean Opt-out: Provide an official configuration toggle in mempalace config to completely disable the auto-save hook for users who prefer to trigger memory storage manually (e.g., using mempalace mine --mode convos later).
Thanks for considering this!