This is the “make it automatic” setup for Claude Code.

You will get:

• automatic memory recall injected when you submit a prompt
• automatic capture of Claude’s final answer into your memory DB

By default, injection is high-signal and not spammy:

• it injects once per Claude session (so it does not slow you down)
• it prefers memories that include a prior prompt plus a “what worked” answer

How it works

Claude Code hooks run a command and pass JSON on stdin.

• UserPromptSubmit hook:
  • stores your prompt into the memory DB
  • prints a short “Memory lessons” block to stdout
  • Claude Code injects that stdout into context (so Claude sees it before responding)
• Stop hook:
  • reads the transcript file
  • stores the latest assistant response into the memory DB
  • prints nothing

0) Install pastchats-memory so hooks can run it

Hooks need pastchats-memory available in your PATH.

Option A (recommended): pipx install (global command)

pipx install git+https://github.com/Hulupeep/pastchats-memory.git

Option B: use a repo checkout + venv

Clone once:

git clone https://github.com/Hulupeep/pastchats-memory.git ~/pastchats-memory
cd ~/pastchats-memory
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Then reference the full path to the binary in your hook command.

1) Create your memory DB

Pick one global DB path. Example:

pastchats-memory init --db ~/.claude/prompt_memory.db

2) Configure hooks

Edit your Claude Code settings file.

• user-level: ~/.claude/settings.json
• project-level: .claude/settings.json

Add:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "pastchats-memory hook-user-prompt-submit --db ~/.claude/prompt_memory.db --limit 3"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "pastchats-memory hook-stop --db ~/.claude/prompt_memory.db"
          }
        ]
      }
    ]
  }
}

3) Verify

1. Start a new Claude Code session.
2. Submit a prompt.
3. You should see a short injected block starting with:

Memory lessons (from past chats):

Performance notes (so it doesn’t slow you down)

• By default this injects once per session, and stores one row per stop.
• Use local embeddings (default) for offline speed.
• Keep --limit 3.
• If you have a huge DB and it feels slow, enable sqlite-vec later.

Tuning

Inject once per session (default)

pastchats-memory hook-user-prompt-submit --db ~/.claude/prompt_memory.db --limit 3 --policy once-per-session

Inject at most every 15 minutes (cooldown)

pastchats-memory hook-user-prompt-submit --db ~/.claude/prompt_memory.db --limit 3 --policy cooldown --cooldown-seconds 900

Ignore tiny prompts

pastchats-memory hook-user-prompt-submit --db ~/.claude/prompt_memory.db --min-chars 40

If you want true MCP usage

You can run the MCP server (pastchats-memory-mcp) and have Claude call recall() as a tool.

For hooks, the CLI approach above is simpler and faster.