dsh-memory-gate

When building long-term conversations, retrieving a memory doesn't guarantee it should be injected—how can we ensure relevant memories don't mislead the model, contaminate context, or leak sensitive information? dsh-memory-gate is a long-term memory plugin for DeepSeek Harness designed to address this challenge. It employs a mechanism called CBDC (Claim → Belief → Decision → Consumption) to adjudicate every recalled memory (use/verify/ignore) before injection, thereby safeguarding context integrity while maintaining relevance.

Its core capability lies in a closed-loop process of "adjudicate first, inject later, learn continuously": memories are assigned confidence scores upon storage and undergo four-stage CBDC evaluation before each use. Post-use feedback (e.g., helpful/harmful) dynamically adjusts weights, making memory more accurate over time. The plugin also features near-duplicate merging, dual-channel recall (capsule + trigger words), three runtime modes (shadow/assist/enforce), self-diagnostics based on health metrics, and default limits on injected quantity and length to control costs—all audited in a local SQLite database.

If you're a DeepSeek Harness user seeking stable, controllable long-term memory capabilities—especially when you need to mitigate risks from memory injection, ensure transparency and traceability in memory usage, or manage complex memory scenarios across multiple scopes (session/workspace/global)—dsh-memory-gate is a valuable tool. It is particularly suited for developers, researchers, or teams prioritizing safety and control, helping you build more reliable AI assistant experiences in long-term interactions.