Self-evolve memory plugin for DeepSeek Harness

Key features

• Three-layer memory (L2/L1/L0)The original log never enters prompt words → structured files → the summary injected at each step, and the fixed cost does not increase with memory.
• search(topic) — Retrieve user preferences by task topic
• The plugin only contains a "learning mechanism" and does not contain any personal data - each user's profile is learned independently from their own historical sessions.
• No-sense injectionL0 injects system prompt words at each step according to the token budget (default 600 characters), and automatically rebuilds with session compression
• Relevance retrievalThe memory tool selects the Top-K based on the current task topic and discards it after searching, leaving no context.