dsh-self-memory

Every fresh DeepSeek Harness session starts from zero. The pitfall you patched half an hour ago, the solution you validated in the previous window—all gone. dsh-self-memory closes that memory gap by structurally recording problems, multiple solutions, script paths, and full model-call traces to local disk. The instant a new session opens, the plugin auto-retrieves relevant history and injects it as a source-attributed user message, so the agent checks before it acts.

The loop runs without manual intervention: hit a problem, read the injected history, try what fits; mark a hit on success, log the root cause on failure; if nothing matches, solve fresh and write the result back. The retrieval engine supports precondition-aware ranking (appliesTo, conflictsWith, and status tags let solutions be annotated as applicable, contradictory, or deprecated), an inverted index that narrows candidates once records exceed one hundred, and generalization matching that swaps version numbers, paths, and hashes for placeholders before scoring. Multiple solutions can hang off one problem and be filtered by preconditions. External playbooks can be bulk-imported from files or inline JSON.

The plugin has zero runtime dependencies, uses only Node built-in modules, keeps every byte under a local directory, and works fully offline so no conversation ever leaves the machine. It is built for heavy DSH users who refuse to re-solve the same puzzle every session, and for engineering teams that want to hand their accumulated runbooks straight to the agent and have them taken up where they left off.