SodaMem/dsh-plugin-sodamem

README excerpt

dsh-plugin-sodamem A standalone plugin for the DeepSeek Harness ( dsh ) that wires in SodaMem as a persistent memory layer. New here? SodaMem is an open-source long-term memory store for LLM agents: you ingest conversation turns into it and it extracts durable facts, then serves a prompt-ready evidence block back on demand. It runs as a local daemon. This repo is only the dsh plugin — the memory engine itself lives at github.com/SodaMem/SodaMem, and you need it running for this plugin to do anything. The plugin installs SodaMem as a memory layer , not as a tool in the model's tool bag. - Recall — before every turn, the plugin fetches a prompt-ready evidence block from SodaMem and contributes it to the system prompt. - Retain — when every turn closes, the plugin ingests that turn's messages back into SodaMem. Neither one is a tool call, so neither one depends on the model deciding to use it. ⚠️ Not ready for use: recall lands one turn late Verified against a real dsh runtime and a real SodaMem daemon ( npm run test:integration ): recall does not reach the model on the turn that asked. AgentLoop.preStep() assembles the system prompt and projects the runtime-context snapshot before it…

Category