DeepSeek Harness Long-Term Memory Plugin: dsh-memory-amem

`dsh-memory-amem` is a plugin for DeepSeek Harness (DSH) designed to inject A-MEM-style long-term memory. The plugin captures user messages through a six-step pipeline—filtering, analysis, retrieval, evolution, application, and persistence—associates new information with prior memories using hybrid retrieval and LLM-based decision-making, and embeds the results into the system prompt to aid model recall. It also provides security defense mechanisms and tools such as search. On the LoCoMo benchmark, the plugin significantly improves multi-hop reasoning, enabling DeepSeek models to outperform GPT-4 in closed-book question answering.