dsh-honcho-memory

Honcho v3 long-term memory plugin for DeepSeek Harness (DSH). Ordinary conversations are written into Honcho automatically; before each model request the plugin retrieves session summary, user representation, peer card and related conclusions for the current question and injects them through DSH's dynamic runtime context.

• Automatically writes real user messages and model-visible replies to Honcho; plugin context, tool results and reasoning chains are not written.
• Each DSH session maps to its own dsh-<session-id> Honcho session by default; a fixed shared session can be configured instead.
• Two peers by default: Charles (user) and deepseek (assistant); the assistant observes the user and forms its own representation.
• First-turn auto recall: waits for bounded Honcho requests during DSH's system-prompt/assemble stage so async results are ready for the current turn.
• Shared-memory cold start: falls back to Honcho's omniscient user representation when the deepseek -> Charles profile has not formed yet.
• Tools: memory_store, memory_search, memory_context, memory_reason (Honcho dialectic cross-session reasoning) and memory_status.
• Configurable baseUrl, apiKey, workspace, peers, sessionId/sessionPrefix, autoCapture, autoContext, context chars/tokens/timeout, searchScope and conclusion limi
• Failures are written to DSH logs; memory_status shows queue, auto-write count, context load count and recent errors.