dsh-living-memory

Most agent-memory plugins solve one problem: write it down, search it up. But memory does not tidy itself. Duplicates pile up, stale todos float to the top forever, and when a new fact contradicts an old one the store silently overwrites history. What you get is not memory—it is an ever-growing log.

dsh-living-memory treats memory as a living system. A nightly patrol automatically deduplicates, merges, applies temporal decay, detects conflicts, and updates the knowledge graph. At query time seven independent signals (full-text, vector, decay, relevancy, co-occurrence, Personalized PageRank, and edge-hop) are fused via RRF; if any channel is missing the ranking degrades gracefully instead of crashing. Full-text search uses FTS5 with jieba tokenization, making Chinese a first-class citizen. Everything lives in a single local SQLite file. No telemetry, no cloud dependency, and the vector path is fully off unless you configure it.

Built for anyone running long-lived agents on DeepSeek Harness who needs Chinese-language memory, wants zero external data dependencies, and expects memory to behave like a hippocampus—forgetting what is stale and resurfacing what is relevant. Fully functional out of the box, with every subsystem independently switchable for debugging.