dsh-agent-memory

Every DeepSeek Harness conversation is stateless by design—the agent remembers nothing between turns, and users find themselves re-explaining context over and over. dsh-agent-memory closes that gap with zero runtime dependencies: it hooks into session/event to capture user messages in real time, stores them under the local memory directory, and then injects per-session memory before the latest user message via agent/pre-step, giving the agent a 2 KB window of prior context without any manual copy-paste.

Beyond injection, the plugin surfaces a live Memory tab inside the Web UI that shows capture status, session index sizes, and preference digests at a glance. Ten read-only data endpoints let external tools query the evidence store without arbitrary-path exposure. On the safety side, every write passes through a redaction layer that strips bearer tokens, cookies, JWTs, private keys, and other secret patterns; all routes enforce the browser-trust fence (fail-closed when unavailable), and a circuit breaker isolates any single capability failure so the DSH main loop is never affected.

It is best suited for DSH users running multi-turn, context-heavy workloads—code review, architecture design, long-form writing—where losing prior context each turn is genuinely costly. The plugin stays out of the global-injection path managed by the official agent-instructions, writes only under its own store root, and survives DSH upgrades with a simple re-install, requiring zero changes to DSH configuration.