memory-system_skill

Overview

This skill provides a local memory system that indexes Markdown files into a SQLite-backed store to enable cross-session semantic search. It supports incremental indexing, hybrid search (vector + full-text), adding and cleaning memories, and program-friendly JSON output. The system is optimized for repeatable, fast lookups without reprocessing unchanged files.

How this skill works

The skill scans a memory directory of .md files, computes content hashes, extracts embeddings, and stores metadata and vectors in a SQLite database for hybrid semantic + keyword search. Commands allow adding new notes, running an incremental index, performing JSON search queries, checking status, and cleaning old entries. Searches read only from the database (not source files) and return structured JSON suitable for programmatic consumption.

When to use it

• You need cross-session recall of user facts, notes, or context.
• You want fast semantic search over Markdown notes without a remote service.
• You need to add single items programmatically or via a conversational trigger.
• You want to periodically clean or audit old memories.
• You need machine-readable JSON search results for automated workflows.

Best practices

• Keep memory files as small, focused Markdown notes to improve relevance and retrieval speed.
• Use the incremental index frequently; unchanged files are skipped via SHA256 hashing to save work.
• Run searches against the SQLite DB (not source files) to get deterministic, fast responses.
• Enable JSON output for downstream automation and use human-readable output for manual review.
• Regularly run status and cleanup to keep the database size manageable and remove stale entries.

Example use cases

• Save a user preference or profile snippet when they say “remember this” and recall it in a later session.
• Index a project’s meeting notes folder and ask semantic queries like “what decisions were made about X?”
• Automate ingestion from other tools by writing .md files into the memory directory and triggering an incremental index.
• Run periodic cleanup to remove notes older than N days and audit memory health with the status command.
• Integrate JSON search responses into a chat agent flow to provide context-aware replies without reading all files at runtime.