dsh-memory-amem

DeepSeek Harness ships multiple chat agents but no native long-term memory: every new session starts from zero and the model has no way to recall prior conversations. dsh-memory-amem brings the A-MEM (NeurIPS 2025) paradigm into DSH, capturing every user message as a structured note and using an LLM-driven evolution step to link new memories to existing ones, so cross-session knowledge no longer disappears.

The plugin runs a six-stage pipeline behind the scenes: a deterministic admission gate filters out noise and sensitive content, an LLM extracts keywords and context tags, BM25 blended with TF-IDF retrieves the most relevant neighbor notes, and a final LLM decision strengthens, updates, or skips the evolution. On every turn the most relevant historical notes are dynamically injected into the system prompt, letting the model perceive prior context without an explicit tool call. Four tools are also exposed for on-demand use: memory_search, memory_add, memory_recent, and memory_stats, so the model can query or extend the knowledge base whenever it needs to.

This plugin is best suited for users who need persistent cross-session dialogue, multi-turn long-horizon reasoning, or agents within DSH that must build a continuously growing knowledge base. On the LoCoMo long-conversation benchmark, even paired with a small DeepSeek model, recall quality exceeds closed-source large-model baselines that have no memory at all, delivering near-long-context retrieval at a fraction of the inference cost.