omdsh-dev/dsh-mnemon

What it solves

dsh-mnemon provides a composable memory system for the DeepSeek Harness (DSH), allowing AI agents to maintain and access information across different timescales and formats without duplicating data. It solves the problem of managing fragmented context by unifying runtime preferences, searchable documents, and long-term evidence into a single view for the agent.

How it works

The system uses a three-tier architecture to organize memory based on how it reaches the agent:

• Runtime: Stores immediate preferences and facts needed for the next turn, projected as compact USER/MEMORY context.
• Documents: Stores designs and procedures, which the agent accesses via search and narrative reading.
• Memory Spaces: Stores durable facts and relationships, retrieving on-demand evidence from various backends (Providers).

This is achieved through a "Source + Strategy $\rightarrow$ View" pipeline. Sources own the data and read/write operations; Strategies determine how those sources participate in a turn (selection and retrieval); and the Core compiles these into an immutable View that the DSH Host pins to the executing turn.

Who it’s for

It is designed for developers using the DeepSeek Harness (DSH) who want to implement sophisticated, tiered memory management for their AI agents, as well as plugin authors who wish to create new memory sources or storage providers.

Highlights

• Three-Tier Memory: Distinct handling for short-term runtime context, medium-term documents, and long-term durable evidence.
• Composable Architecture: Support for custom Sources and Strategies via an extension SDK.
• Extensive Provider Support: Integration with multiple backends including Mem0, RetainDB, and Mnemon Native.
• Unified Interface: Provides a consistent Sidebar view for managing memory across different scopes (global, workspace, and custom).