AI Agent Hub
Back to plugins
🧠

dsh-memory

Memory Updated 2026.08.14

Run the following command in DeepSeek Harness:

dsh plugin install Jesse-njx/dsh-memory

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install Jesse-njx/dsh-memory inside a DeepSeek Harness session to install the plugin; the source is available at https://github.com/Jesse-njx/dsh-memory .

About this plugin

Most agent memory pipelines rely on lossy summarization: once a distiller mangles a detail, the original context is gone and the agent confidently reasons from the corrupted version. dsh-memory takes a different route — it builds no memory store at all. Instead it treats DSH's existing append-only session log as ground truth and layers a thin, citation-backed index on top of it.

After each session, an async distillation pass extracts durable facts — user preferences, project decisions, environment quirks, corrections — into small markdown files under ~/.dsh/memory/. Every fact carries a precise citation (session ID plus event range), and a single memory_expand call restores the exact original log excerpt, losslessly. Files are human-readable, git-diffable, and trivially deletable; every distillation run is recorded in a JSONL audit log. No vector database, no knowledge graph, no auto-injection of full memory bodies.

If you build agents on DeepSeek Harness and want every recalled fact to be traceable back to the specific log events it came from — with a file-system footprint instead of an opaque embedded store — dsh-memory is the lightweight experiment built for that.

Use Cases

  • Retain project decisions, user preferences, and environment quirks across sessions without re-explanation
  • When a recalled fact is off, trace it back to the exact original log events via citation
  • Audit, edit, or delete agent memory with plain git diffs and a local editor

Best For

  • Developers building agents on DeepSeek Harness
  • Engineers who want memory state to be auditable, diffable, and deletable
  • Teams that need cross-session memory without a vector database or knowledge graph