deepseek-harness-mimo-memory

Every new session starts with the Agent knowing nothing about your project context, hardware setup, architecture choices, or coding conventions. It typically burns one or two turns re-reading documents and re-aligning, and a mid-task interruption leaves no trace to resume from. mimo-memory folds persistent cross-session memory directly into DeepSeek Harness: at session start it injects MEMORY.md and checkpoint.md into the system prompt with zero tool calls, so the Agent grasps the full picture on turn one; when the Agent goes idle it triggers a background Dreaming cycle that silently distills rule preferences and verified conclusions from recent dialogue into memory files; and it registers an update_project_memory tool so the Agent can append or replace structured facts by section. All memories are isolated per workspace path, with configurable max injection length and optional auto-distillation.

Best suited for developers who treat DeepSeek Harness as a long-running workspace: juggling multiple projects, needing consistent context across sessions, and wanting the Agent to accumulate knowledge on its own during idle time instead of re-deriving it from scratch. No separate memory scripts to maintain—install once and get a self-evolving project memory at minimal operational overhead.