This is a cloud-based cross-session memory plugin for DeepSeek Harness: structured entries + vector retrieval, persisted in S3-compatible object storage (AWS S3 / MinIO / R2 / Alibaba Cloud OSS).

Introduction

By default, DSH stores memory only in the local instance. If you change your deployment environment or machine, this memory will not be preserved. dsh-memory-s3 persists memory to S3-compatible object storage, enabling memory sharing and continuation across devices and sessions.

Core Features

  1. S3-compatible storage backend: Supports entry CRUD, listObjects synchronization, optimistic concurrency, and custom endpoints (compatible with MinIO / R2 / OSS).
  2. Structured memory model v2.1: Defines five entry types: preference, project, decision, history, and moment; supports metadata such as importance/tags/source.
  3. Vector semantic search: Provides pluggable embedding support, supports cosine top-k retrieval, and hybrid recall using metadata/keywords.
  4. Local cache projection: Supports frozen snapshots, a read-only local cache, asynchronous revalidation from S3, and offline fallback mode.
  5. Write approval gate: All write operations must pass through DSH’s approval seam; they are invisible and unmodifiable by the model.
  6. Three-chain audit: Records approval pairs, the audit ledger, and snapshots, so all writes can be reconstructed from session logs.
  7. Photo/file attachments: Supports attaching files to entries; binary objects are stored in S3 and locally verified.
  8. Backlink index (v2.1): Automatically backfills a local backlink index and supports querying “who referenced this entry”.
  9. Layered snapshot injection: Injects snapshots in the order Bonds → Moments → Facts, with support for associated reference markers.

Installation and Activation

Use the official CLI to install it and add the plugin to the Web profile:

npx @deepseek-ai/dsh plugin --profile web add https://github.com/Re-s/dsh-memory-s3.git

Tool List

After installation, DSH exposes the following tools for memory management:

  • memory_s3_save: Save a memory entry (passes the approval gate).
  • memory_s3_search: Keyword search.
  • memory_s3_recall: Semantic recall.
  • memory_s3_list: List/filter.
  • memory_s3_backlinks: Query backlinks (no approval required; reads the local index).
  • memory_s3_update: Update an entry (passes the approval gate).
  • memory_s3_delete: Delete an entry (passes the approval gate).
  • memory_s3_forget: Suppress automatic injection without deleting.
  • memory_s3_attach: Attach a file (passes the approval gate).
  • memory_s3_get_file: Download an attachment locally (no approval required; validates).
  • memory_s3_detach: Remove an attachment (passes the approval gate).
  • memory_s3_sync: Manually synchronize S3 increments.

Permissions and Notes

  • No GUI form entry: The plugin does not have a visual configuration form; manual configuration is required via settings.yaml. It is visible in the plugin’s Plugin list, but not visible in Plugin configuration.
  • Permission requirements: Running the plugin requires the following permissions: harness:tool, filesystem:read, filesystem:write, network:https, subprocess:none, shell:none, python:none, credentials:env.
  • Version compatibility: Compatible with DSH 0.1.0-rc.6 and 0.1.2-rc.1.
  • Credential management: S3 credentials are read only from environment variables (AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY) and are never written to disk.
  • GitHub: https://github.com/Re-s/dsh-memory-s3
  • Directory page: https://www.skillhub.cn/plugins/Re-s/dsh-memory-s3