dsh-plugins: bounded cross-session memory (MEMORY.md/USER.md) and skill lifecycle curation, ported from hermes-agent

[hyls9527]

Two ecosystem plugins porting the two mechanisms I consider the core of NousResearch/hermes-agent, reimplemented on dsh's own conventions.

Repo: https://github.com/hyls9527/dsh-plugins (tagged dsh-plugin)

@hyls9527/dsh-memory

Bounded, file-backed, cross-session self-maintained memory: $DSH_HOME/memories/MEMORY.md (the agent's own notes) and USER.md (what it knows about the user).

• §-separated entries with character budgets (2200/1375 by default), so files stay human-editable and the model consolidates instead of growing unbounded memory.
• One memory tool: add / replace / remove / batch (atomic multi-op).
• A frozen snapshot is injected as a durable user/message at the first admitted step of each session — mid-session writes never touch it, so the prompt prefix stays stable for the whole session.
• nudgeInterval persistence reminders (default every 10 write-free turns).
• Durability guards: cross-process locks, atomic renames, refusal to overwrite unreadable files, and drift detection (external edits are snapshotted to .bak.<ts> and refused).

Model-visible ⟺ logged is satisfied without new session event vocabulary: the snapshot is a durable message with a memory source, and tool calls are logged by the harness itself.

@hyls9527/dsh-skill-curator

Minimal skill lifecycle curation: every skill tool call is recorded into one usage document, and a skill_curator tool reports usage and staleness (derived from a configurable threshold, default 21 days), archives unused skills out of the catalog, and restores them.

Install

dsh plugin --profile web add @hyls9527/dsh-memory
dsh plugin --profile web add @hyls9527/dsh-skill-curator

Both are dsh bundles (a cordis.patch.yml inserting one row each), built against @deepseek-ai/dsh-* 0.1.0-rc.6, with 100% per-file vitest coverage and verified end-to-end in a real profile (snapshot injection, a live memory write, and usage tracking). Feedback welcome.

[hyls9527]

Update — the learning loop is now complete.

A third plugin landed: @hyls9527/dsh-skill-create — skill authoring.

• skill_create tool with create / update: persists a reusable procedure as a user skill in the exact directory-bundle shape the harness skill provider parses (kebab-case name, name/description/whenToUse frontmatter, size budgets). The registry watches the root, so a landed skill becomes catalog-visible immediately — verified end-to-end: a live agent created a skill, and the very next catalog message in the running harness listed it.
• A creation nudge (default every 10 turns) reminding the model that reusable procedures belong in the skill library, mirroring hermes-agent's creation_nudge_interval.

Together the three plugins form the full self-improving loop: remember (dsh-memory, MEMORY.md/USER.md) → create (dsh-skill-create) → track & curate (dsh-skill-curator). 168 tests, 100% per-file coverage.

dsh plugin --profile web add @hyls9527/dsh-skill-create

[hyls9527]

Archived.

The repository is now archived and no longer maintained. The exercise concluded that these mechanisms are not useful for DeepSeek Harness: the harness deliberately keeps session state in the session log and treats skills as user-authored, vetted inputs, and the official repo has repeatedly, in writing, chosen not to build them. The code remains available for reference; thanks to anyone who read along.