dsh-plugin

Memory Lake for DeepSeek Harness (dsh): persistent, cross-session, cross-project, cross-device memory.

dsh ships no native memory capability, so this plugin is not a bridge to one — it IS the harness's memory layer. It contributes:

• Three tools — memory_search, memory_remember, memory_forget — backed by the memorylake CLI (never by direct HTTP, never through a shell).
• A prompt section owning the policy: when to remember, how to phrase recall queries, how to read results, and that a failed search is never an empty one.
• A one-line status context reporting connectivity — most importantly when the backend is UNREACHABLE, because a silently unavailable memory backend is indistinguishable from "you never told me that".
• Two user-invocable skills — type /memorylake-init for the guided setup wizard (CLI download with mandatory checksum verification, login, shared config) and /memorylake-status for the end-to-end health check. The model cannot trigger either.

Install

dsh plugin --profile web add @memorylake/dsh-plugin
dsh --profile web --dump-config   # shows the "@memorylake/dsh-plugin" layer
dsh web

web is dsh's built-in Web UI profile (auto-initialized on first use); any other profile name works the same way — the plugin composes with whatever surface the profile mounts.

The package publishes with built lib/, so no build allowance is needed. A github: install also works — a self-contained prepare script builds from source — but pnpm ≥10 will ask you to allowlist the build; prefer the registry install.

Zero-config on a machine that already has Memory Lake

Identity and switches live in the shared ~/.memorylake/ tree, not in this plugin:

• ~/.memorylake/credentials.toml — the CLI's login state
• ~/.memorylake/harness/config.md — workspace, actor, and feature flags
• <repo>/.claude/memorylake.local.md — optional per-project override, merged key by key over the global file (a one-line sync_on_write: false override keeps the global workspace working)

If the Claude Code or Codex integration already wrote these, a dsh session picks them up with zero additional steps: status line, tools, recall — all live. Config files are re-read on every call, so /memorylake-init finishing mid-session takes effect on the very next message, no restart.

When nothing is configured, the plugin is completely silent: no network requests, no prompt injection. Only an explicitly invoked memory tool answers, and only to point at /memorylake-init.

Shared config keys (config.md / memorylake.local.md frontmatter)

Key	Default	Meaning
workspace	—	Memory Lake workspace id. The only required key; absent means the plugin stays silent
actor	—	Actor id facts are attributed to; required for writes
enabled	true	false switches the plugin off entirely
sync_on_write	true	false makes memory read-only (the canonical per-project override)
status_line	true	false suppresses the session status line

Keys consumed by the other harnesses (remind_on_read, sync_deny, projects, project_custom_id) are parsed and preserved but not consumed by this plugin's v1.

Deployment knobs (cordis config)

Only operational tuning lives in the bundle's rows; override by id in your profile's cordis.patch.yml (a patch replaces the row's whole config).

Row memorylake (@memorylake/dsh-plugin/service):

Key	Default	Meaning
binaryPath	—	Absolute path of the memorylake binary, overriding resolution (PATH, then ~/.memorylake/bin/memorylake)
timeoutMs	30000	Hard deadline per CLI invocation
killGraceMs	2000	SIGTERM→SIGKILL grace on timeout/cancel
maxOutputBytes	1000000	In-memory cap per collected CLI stream

Row memorylake-tools (@memorylake/dsh-plugin/tools):

Key	Default	Meaning
topKMax	10	Upper bound for the model's top_k
statusTtlSeconds	600	Status-line refresh interval and cache TTL

Behavior notes

• Authentication is CLI login state. The dsh subprocess seam scrubs credential-shaped environment names from children, so an ambient MEMORYLAKE_API_KEY is never forwarded — by design, and matching the CLI's own "env vars alone are not a session" rule. Log in once with memorylake auth login (the init skill walks through it).
• Payload before exit code. The CLI prints its full JSON payload and then encodes the business outcome in the exit code (fact delete exits non-zero when ids were not found). The plugin parses first and classifies second.
• Failure is not emptiness. Unavailable backends produce an explicit notice in the tool value and a loud status line, instructing the model to attribute empty results to the connection rather than to missing memory.
• Scores are ordering-only. Relevance scores sort results and are then discarded; they never reach the model (they are weakly calibrated, and a model shown a number treats it as authority).
• No auto-recall, no session memory digest. Both were evaluated and rejected in the Claude Code integration: fixed token cost on sessions that never touch memory, no behavioral gain. The status line reports connectivity only.

v2 roadmap (deliberately not in v1)

• Session→conversation cook: append dsh session transcripts as Memory Lake conversations and let the backend distill them into memories — dsh's "model-visible ⟺ logged" invariant makes the transcript complete, which no other harness guarantees. Privacy defaults off.
• Document search: Library upload + project import, and passing --projects on search (without it the server returns zero document hits — facts are unaffected, which is why v1 is facts-only).
• Project identity: the explicit-custom-id → normalized-git-remote → physical-path rule shared with the other harnesses.
• Session-start facts digest: a small, config-gated summary — dsh has no local MEMORY.md, so unlike the other harnesses this may carry real value, but it must first justify its token cost.
• Upstreaming as an in-tree packages/memorylake/ capability seam.

Development

pnpm install
pnpm test        # vitest: unit + integration + real-Loader composition
pnpm build       # tsc → lib/

Tests run against a mock memorylake binary and an isolated data tree; no network, no API key, and no touching your real ~/.memorylake.