dsh-memory

Persistent cross-session memory plugin for DeepSeek Harness. Facts the model saves are stored durably per workspace (the owning session's cwd) on the harness storage-domain seam, and injected as a bounded prompt section into later sessions in the same workspace — so something important established in one session carries over to the next.

• ctx.memory service: save / list / search / forget
• Model tools: memory_write, memory_list, memory_search, memory_forget
• Prompt section: app:memory (order 50, default 8 lines) in every session of the workspace

Requirements

The harness profile must mount the storage seam with a backend:

• @deepseek-ai/dsh-storage (the hub)
• @deepseek-ai/dsh-storage-json (default, human-readable JSON files) or @deepseek-ai/dsh-storage-sqlite
• @deepseek-ai/dsh-storage-domain (the domain facility this plugin opens the memory domain on)

The shipped web profile already mounts all three with the json backend — no extra setup needed there.

Install

# from this repo (sources build on install via prepare)
dsh plugin --profile web add github:guntur-d/dsh-memory

# or from npm once published
dsh plugin --profile web add @guntur-d/dsh-memory

The dsh.bundle.patch manifest field makes the plugin a bundle layer: the included cordis.patch.yml row activates the plugin automatically in the profile. GitHub-topic discovery: dsh-plugin.

Usage

The model calls memory_write when something durable is established (a port, a preference, a non-obvious project fact). You can prompt it explicitly:

Please remember: the staging server listens on port 3080 and the deploy script is deploy.sh

A later session in the same workspace starts with an app:memory section listing the recent memories, and can memory_list / memory_search / memory_forget (by id) on demand.

Config

interface Config {
  maxEntriesPerWorkspace: number // default 500; the oldest are evicted past it
  maxTextChars: number           // default 2000; longer saves are refused
  maxContextEntries: number      // default 8; memories rendered into the prompt section
  maxListEntries: number         // default 50; memories one list/search call returns
}

The storage backend route is decided by the deployment's dsh-storage-domain config (routes for the memory domain, else its default backend).

Semantics

• Entries are scoped by the owning session's cwd; sessions without a cwd share one anonymous workspace.
• A monotonic seq (minted on the domain's atomic write chain) keeps ordering and eviction deterministic even when wall-clock timestamps tie.
• Saves and forgets await durability on the backend; reads are synchronous from the domain's in-memory state.
• The model decides what is worth saving — no automatic summarization.

Develop

pnpm run check       # typecheck against a deepseek-harness checkout (tsconfig.check.json)

Notes:

• The npm-published @deepseek-ai/* dependency chain is currently incomplete (a peer package @deepseek-ai/dsh-type-meta is not published), so a from-npm pnpm install of this plugin fails until upstream completes the publish chain. The tsconfig.check.json typechecks this source against a local deepseek-harness checkout instead (edit the absolute path in its paths if your checkout lives elsewhere).
• Until the npm chain is fixed, install the plugin from a checkout: dsh plugin --profile web add . from this repo (the profile resolves @deepseek-ai/* through the harness installation).