dsh-noema

Every AI coding agent starts each conversation with amnesia. You spend the first twenty minutes re-explaining your project conventions, decisions, and constraints — only for the next session to forget them all. DSH Noema fixes this by connecting DeepSeek Harness to Noema, a local-first, non-vector memory system. Instead of fuzzy vector embeddings, your agent's memories are stored as plain, inspectable Markdown files under ~/.agent-memory/, and a noema_recall tool loads the relevant context at the start of every session. You can search, browse, audit, and even ask the agent to explain why a memory was or wasn't recalled.

Where it really stands out is migration. If you've been using Codex, Claude Code, opencode, Cursor, Grok, WorkBuddy, Antigravity, Trae, Qoder, or Hermes, the noema_import tool reads those tools' memory files (AGENTS.md, CLAUDE.md, MEMORY.md, rules, and more), splits them into sections, and imports each one as a durable memory — with a content-keyed ledger that deduplicates across runs and across tools that share the same files. A settings page gives you full control: server command, memory root, token budgets, timeouts, and a memory browser for adding, reviewing, and deleting entries directly.

This plugin is built for people who actually run an agent day in and day out. The idle timeout defaults to never, a keep-alive loop restarts the memory server in the background if it crashes, and after the first boot everything hot-reloads — rebuild the host plugin and the running DSH reloads it without a server restart; rebuild the client bundle and the browser hot-swaps without a page refresh. If you juggle multiple AI coding tools and want to consolidate all that accumulated context into one durable, inspectable, local-first memory for DeepSeek Harness, DSH Noema is for you.