memory-cn

What This Skill Does

The memory-cn skill is a highly specialized optimization module for OpenClaw users working with Ollama-based Chinese language environments. It addresses the critical FTS5 unicode61 tokenization bug, where the database treats continuous Chinese characters as a single, indivisible token, effectively breaking full-text search capabilities for Chinese queries. By implementing a systematic approach—including structured logging, strategic keyword tagging, and weighted hybrid search parameters—this skill restores semantic search accuracy and elevates memory retrieval hit rates from a baseline of 55% to near 100%.

Installation

To install this skill, use the following command in your terminal: clawhub install openclaw/skills/skills/abczsl520/memory-cn Once installed, ensure you execute openclaw memory index --force to initialize the optimized indexing structure.

Use Cases

This skill is essential for power users maintaining long-term conversational memory with Ollama models. Use it when:

1. Your search queries in Chinese are consistently returning 'no results' despite the information being present in your history.
2. You are managing large knowledge bases (lessons) or project files that have grown beyond 8KB per file.
3. You need to transition from a disorganized log structure to a highly queryable, tag-based memory retrieval system.
4. You are using small-scale local LLMs (e.g., 0.6B models) and need optimized chunking to improve context relevance.

Example Prompts

1. "OpenClaw, please run the diagnosis script and report on the current health of my memory file tags and FTS5 indexing."
2. "I'm having trouble retrieving old project details in Chinese; can you patch my memory search settings to the optimized hybrid configuration?"
3. "Run a maintenance cycle: add missing tags to my project logs, compress files exceeding 8KB, and force a rebuild of the memory index."

Tips & Limitations

The memory-cn skill relies on space-separated Chinese keywords to bypass FTS5 limitations. Users should ensure that their manual memory entries follow the <!-- tags: ... --> header convention to maximize indexing performance. While the 0.75 vector weight is highly effective for smaller local models, you may need to adjust the minScore parameter if your specific LLM exhibits unexpectedly low cosine similarity values. Always backup your memory/ directory before performing bulk compression tasks. This skill is specifically tuned for local storage deployments and may require additional permissions if your memory store resides on a network-mounted drive.