Culeot/dsh-agent-memory

README excerpt

dsh-agent-memory Cross-session long-term memory plugin for DeepSeek Harness (DSH). 中文文档 What it does Remembers facts, preferences, decisions, and lessons across sessions, so a new session starts with what earlier sessions already learned. Storage lives in $DSH HOME/storages/memory.json (plain JSON — inspectable, git-friendly). No MCP servers, no vector database, no embedding API, no extra runtime dependencies. Features - Cross-session persistence — write in session A, recall in session B; new sessions pick up where old ones left off. - Relevant memories injected on demand — before each model step, the plugin recalls the most relevant memories for the user's current message (via DSH's agent/pre-step hook) and injects up to injectCount short summaries (default 3). You say "continue the math workbook project" → the project's decisions and lessons appear automatically. No every-turn broadcast of a fixed hot set — you only pay for memories that match what you just said. - Self-correction loop — when the same error (same code/message) fires repeatedly ( lessonizeAfter , default 2), the plugin nudges the agent to write it as an importance-3 lesson; the lesson then auto-injects on related …

Category