dsh-mnemon

Memory systems often face an awkward trade-off: either cram every kind of knowledge into a single database, or let context balloon out of control on every turn. dsh-mnemon solves this with a three-tier architecture. The Runtime tier projects compact preferences, collaboration rules, and project conventions into every turn. The Documents tier preserves complete narratives like designs, postmortems, and handoffs. The Memory Spaces tier retrieves durable cross-session evidence on demand from nine long-term providers. Each tier has a distinct job, and the first two are run by a deterministic Host that costs no extra model context.

What really sets dsh-mnemon apart is its Agent-driven, supervised workflow. Distilled writes via Remember, bounded evidence retrieval via Agent query, smart routing, and AI metadata generation are all handled by separate background task Agents that never touch your main conversation history. You can assign them a dedicated provider and model, or keep Search and Turn memory as pure deterministic Host reads. The nine providers — from the native local-first Mnemon engine to OpenViking, Honcho, Mem0, and others — keep their capability differences visible; the plugin never invents graph edges, deletion semantics, or enumerable content an engine doesn't actually provide.

It's built for DeepSeek Harness users who want one memory control plane across workspaces and Agents without forcing everything into one database, and for teams who already rely on an external memory service and want to plug it into DSH while keeping their existing workflow. Settings owns reusable provider services, Memory Spaces owns concrete instances and scope, and the WebUI, conversation surface, and Headless mode share the same system. Security boundaries are explicit: provider credentials stay at mode 0600, host calls use argument arrays with shells disabled, and uninstalling the plugin never deletes local or remote memory data.