dsh-hindsight-memory

Forgetting between sessions is a persistent pain point for conversational agents: a fact the user stated last turn is gone the next turn. The usual fix is to have the model invoke a memory tool explicitly, but that costs extra tokens, adds latency, and the model can simply forget to call it.

dsh-hindsight-memory takes a different approach: it wires Hindsight into DeepSeek Harness as pipeline-level long-term memory without exposing any agent-facing tool. Before the first LLM call of each turn, the plugin automatically recalls context using the latest user message as the query and injects the result as a plugin-sourced user message. After the turn ends, it asynchronously retains the User/Assistant exchange. The whole cycle is invisible to the model. Turn-level replay guards prevent redundant recalls across steps, and trivial chit-chat or slash commands are filtered out to avoid wasteful lookups.

Failures are strictly isolated: a failed recall simply skips injection for that step, a failed retain is logged without interrupting the chat loop, and sub-agents or forked sessions participate by default. A dedicated settings section in the browser sidebar lets you edit the API endpoint, key, memory bank, and recall budget in real time with no restart needed. Ideal for developers and teams already running DeepSeek Harness with a Hindsight deployment who want seamless cross-session memory.