dsh-plugin-memory

English | 中文

Session distillation memory for DeepSeek Harness: a dsh-plugin bundle that distills completed turns into durable knowledge entries and surfaces each entry to the model as a memory-<slug> skill. The mechanism ports a prior task-distillation design (distill on completion, merge_with-driven fusion, one Markdown entry per memory plus an index).

What it does

• On every turn/end with reason completed, collects the turn's direct human question, final assistant answer, and tool calls. Answers whose tail asks the user for more input are skipped.
• One auxiliary model call distills the question/answer into strict JSON — title, answer, sources, and optionally merge_with, the slug of an existing entry the model judges same-topic. A reported merge_with triggers a second fusion call over the existing entry plus the fresh distill; a failed fusion keeps the fresh distill but still supersedes the old entry.
• Entries land in the memory directory (default <DSH_HOME>/memory): one <slug>.md Markdown file per entry plus a _index.json mapping slugs to records. Index writes are atomic (temp file + rename); entry writes serialize through one promise chain. Slugs are lowercase ASCII [a-z0-9-] (CJK titles fall back to a 12-hex SHA-1 digest), and model-reported slugs are validated before any path join.
• Every completed auxiliary call appends a log-only memory/distill-call session event (purpose, route, output cap, complete output blocks, optional usage), and every committed write appends a log-only memory/distilled event.
• When a ctx.skills service is mounted, the plugin registers the memory skill provider: one memory-<slug> candidate per entry at rank 550, below every hand-authored skill root. Loading a candidate returns the entry Markdown body as the skill content.

Requirements

The plugin uses vocabulary shipped by the DeepSeek Harness memory change (purpose: 'memory' in dsh-llm and the memory/* session events): it requires a Harness build that includes that change. Older builds still load the plugin and distill correctly; only the DeepSeek thinking-disable mapping and the memory/distill-call replay vocabulary are lost.

Install

From npm (prebuilt, no build permission needed):

dsh plugin --profile <name> add dsh-plugin-memory

From GitHub (sources; pnpm runs the package's prepare build after install):

dsh plugin --profile <name> add github:<owner>/dsh-plugin-memory

pnpm ≥10 requires allowlisting the git dependency's build script; copy the exact key dsh prints into the profile's pnpm-workspace.yaml:

allowBuilds:
  dsh-plugin-memory: true

From a local tarball:

pnpm pack
dsh plugin --profile <name> add ./dsh-plugin-memory-0.1.0.tgz

Then boot: dsh --profile <name>.

Config

Override the row by id: memory in your profile's cordis.patch.yml, e.g.:

- id: memory
  name: dsh-plugin-memory
  config:
    autoDistill: false

Service API

The service registers as ctx.memory:

• list(): Promise<MemoryIndex> — the current index, or an empty one when absent or corrupt.
• readEntry(slug: string): Promise<string | undefined> — one entry's Markdown body.
• settled(session: Session): Promise<void> — resolves once every distillation pending for the session settles.

The optional invariant companion (dsh-plugin-memory/invariant) validates the memory/* event payloads through ctx.invariants; harness builds that know this plugin may mount it as a separate row.

Known Limitations and Deferred Work

• No content-safety policy equivalent to the reference implementation's pre-write filter; distilled answers are stored as produced.
• The incomplete-answer filter is marker-based and can both miss and misfire.
• One memory directory serves every session; per-workspace or per-agent scoping is deferred.
• The catalog lists every entry; there is no per-session cap or size-based eviction.

Publishing

The GitHub repository should carry the dsh-plugin topic so marketplaces and aggregators discover it.

License

MIT — [LICENSE](LICENSE). The plugin ports code from the MIT-licensed deepseek-harness repository.