Introduction

When handling coding tasks, DSH (DeepSeek Harness) often encounters issues such as the AI forgetting previously rejected implementation approaches, repeatedly hitting the same pitfalls, and losing project engineering constraints. The dsh-plugin-memos-code-retrospect plugin replicates mneme’s coding conversation retrospective distillation logic. It addresses the above issues by extracting rejected solutions, pitfall records, and engineering constraints. It does not introduce the mneme binary or create a separate database; instead, it fully reuses the MemOS memory foundation for data persistence.

What Is It

This plugin is maintained by ai-fu-cn and is an MIT-licensed open-source plugin. It extracts rejected solutions, pitfall records, and engineering constraints at the end of a conversation (turn/end), tags them with type:rejected_solution, and writes them to MemOS. Before reasoning for a coding task (agent/pre-step), it performs weighted recall of these memories and injects them into the context, enabling the AI to avoid historical traps.

Core Features

  • turn/end distillation: After a round of interaction ends, the plugin assembles the complete conversation context from session.events (with a configurable truncation length) and calls an LLM to extract three entity types: rejected_solution (rejected solution), pitfall (pitfall record), and constraint (engineering constraint).
  • agent/pre-step weighted recall: Before agent reasoning, the plugin determines the task type and retrieves type:rejected_solution memories only for coding tasks. It boosts their weights (default 2x) and then injects them into the prompt.
  • v2 robust version: Includes fallback retry for LLM JSON parsing (no more than once), similarity deduplication before writing (threshold 0.82), low-value content filtering, and 14/14 unit tests passing.

Installation and Enablement

  1. Prerequisites: Ensure DSH (web profile) and the MemOS local plugin (@memtensor/memos-local-plugin) are installed, and that the Node.js version is >= 18.
  2. Install the plugin:
    dsh plugin --profile web install https://github.com/<your-github-name>/dsh-plugin-memos-code-retrospect
  1. Register it in DSH: Append the following configuration to <profile>/cordis.patch.yml:
    - insert:
        - id: dsh-plugin-memos-code-retrospect
          name: './plugins/dsh-plugin-memos-code-retrospect/dist/index.js'
          config:
            enabled: true
            profileId: default
            recallEnabled: true
            captureEnabled: true
            boostFactor: 2
            recallTopK: 6
            contextMaxChars: 3000
            maxDistillContextChars: 8192
            distillMaxTokens: 2048
            distillRetries: 1
            dedupeEnabled: true
            dedupeThreshold: 0.82
            valueFilterEnabled: true
            sync_to_tencentdb: false
  1. Restart the service:
    Stop-Process -Name dsh -Force
    dsh web --port 3080

After startup, the plugin loads successfully if a `[retrospect]` prefix appears in the logs.

Typical Usage

  1. In a coding task, if errors occur or solutions are rejected, [retrospect] distilled trace=... appears in the logs after turn/end.
  2. Search for tags: type:rejected_solution in the MemOS viewer to view the new trace.
  3. Start a related coding task in the next round; injected rejected_solution recall appears in the pre-step logs, and the model context contains <retrospect_context>.

Notes

  • Distillation quality: Distillation quality depends on the LLM’s instruction-following capability. It is recommended that distillMaxTokens >= 2048.
  • Memory latency: Memory writes occur at turn/end, so memories written in the current round can only be recalled in the next round.
  • Task filtering: For non-coding tasks, type:rejected_solution memories are not injected by default to prevent noise pollution (codingKeywords is configurable).
  • Extension placeholder: sync_to_tencentdb is a reserved extension and is not implemented in this release.
  • MemOS dependency: If MemOS is unavailable, the plugin fails open (it can still load, but memory read/write does not work).
  • Data isolation: The plugin only performs logical processing. All persistence operations call MemOS interfaces, and it does not store data internally. Launching mneme child processes is prohibited.