memory-mine

When to use

Use this skill when the user wants to:

• Index/mine project source files into the palace memory
• Mine conversation history (JSONL, text) into the palace
• Rebuild or refresh the palace knowledge base from disk
• Bootstrap memory for a new project

What it does

Runs pi memory mine to walk a directory and ingest files as palace drawers:

1. Source mining (default): Walks the directory tree, chunks source files (1500 chars, 200 overlap), assigns rooms via mempalace.yaml patterns or directory structure, generates embeddings, deduplicates (cosine >= 0.9), and stores as drawers.

2. Conversation mining (--convos): Parses JSONL (pi-go or Claude Code format) and plaintext conversations, pairs user+assistant turns, and stores as drawers with importance=4.

Steps

1. Check if palace.db exists (run pi memory init . if not)
2. Check if the embedding model is downloaded (run pi memory model download if not)
3. Run the mine command:

# Mine source files from current directory
pi memory mine .

# Mine source files with explicit wing name
pi memory mine . --wing my-project

# Mine conversation files
pi memory mine ./conversations --convos

1. Show the results (processed, added, skipped duplicates, errors)
2. Optionally run pi memory status to confirm the updated state

Arguments

The user can pass arguments after /memory-mine:

• A directory path (default: . current directory)
• --wing <name> to override the wing name
• --convos to mine conversations instead of source files

Examples:

• /memory-mine — mine current directory
• /memory-mine ./src — mine the src directory
• /memory-mine ./chats --convos — mine conversation files

Prerequisites

• pi binary must be built and available (go build -o pi ./cmd/pi)
• Palace DB must be initialized (pi memory init .)
• For semantic search, embedding model must be downloaded (pi memory model download)

Notes

• Respects .gitignore patterns
• Skips hidden dirs, node_modules, vendor, .git, build, target
• Skips files > 512 KB
• Uses mempalace.yaml for room assignment if present
• Duplicates (cosine similarity >= 0.9) are automatically skipped