dsh-self-memory: DSH Self-Memory Plugin
The `dsh-self-memory` plugin is designed to address the lack of cross-session experience reuse for DeepSeek Harness agents. By recording session trajectories and solutions, it provides offline retrieval and closed-loop write-back capabilities, improving the efficiency with which agents solve recurring problems. The plugin offers five core capabilities: multi-keyword association, generalized summarization, fast retrieval, premise filtering, and synonym normalization. It supports hot-mount installation, with automatic collection and injection of historical solutions enabled by default. Users should follow a closed-loop workflow: “check records first → try a match if found → record the root cause on success or failure → develop a new solution if none exists.” The tools include status queries, solution retrieval, record appending, and success/failure marking. Configuration is flexible, allowing the recall budget and storage directory to be adjusted, and it has no external dependencies. Retrieval currently uses keywords and an inverted index; migration to SQLite is recommended when the record volume becomes large.
Read Moredsh-self-memory: A Local Offline Closed Loop for DSH Experience Memory
`dsh-self-memory` is a zero-dependency offline plugin developed for DeepSeek Harness (DSH), designed to address the problem of agents repeatedly explaining background context during debugging. It enables the automatic accumulation and reuse of experience by structurally recording issues, solutions, script paths, and model calls during sessions. The plugin provides core capabilities such as automatic retrieval and injection, external import, premise-aware retrieval, and generalized normalization. Users can retrieve history, record new solutions, mark hits or failures, and import data via the command line. Currently, it only supports keyword search, and storage performance may be limited with tens of thousands of records, but it can effectively reduce the cost of repetitive work for agents.
Read More