dsh-memory-manager

AI Agents repeatedly lose context in long conversations and cannot reuse knowledge across sessions, while most memory solutions depend on cloud embedding services, making them impractical in offline or sensitive environments. dsh-memory-manager addresses this with a local-first, zero-cloud, zero-native-dependency memory stack: a per-session sliding window handles short-term context with automatic truncation and summarization; long-term recall fuses cosine-similarity vector search with keyword term-overlap scoring, and gracefully degrades to hash-vector plus keyword blending when no embedding model is configured, so it works out of the box. On top of that, a four-layer LLM distillation pipeline (L0 raw dialogue to L1 atomic memories to L2 scenes to L3 persona) is governed by TTL expiry, importance scoring, storage-cap LRU cleanup, and per-session gating, keeping memory both rich and bounded. Security and collaboration are first-class: AES-256-GCM encrypted persistence, session isolation with optional global sharing, versioned JSON/JSONL import-export, and a built-in web console covering browsing, semantic search, statistics, logs, and model downloads, all accessible via a DSH settings sidebar embed. It is well suited to developers building multi-turn dialogue agents in the DeepSeek Harness ecosystem, whether prototyping locally offline or turning conversations into a searchable, distillable, portable knowledge base for the whole team.