@luisarg/memory-auto

dsh plugin --profile web add @luisarg/memory-auto

@luisarg/memory-auto

Cordis bundle that auto-captures session memory for

DeepSeek Harness (DSH):

it watches sessions and digests transcripts into an OKF vault at checkpoints —

idle, git commit, compaction, and session end.

The digest runs in-process: extraction uses the harness's own LLM service

(ctx.llm.stream, provider/model configurable, deepseek-official by

default), and the resulting OKF entries are written through the vault's MCP

server (store_* tools over stdio), keeping the Markdown source of truth and

the SQLite FTS5 index in sync. No external CLI and no stored credentials — the

plugin uses the same key DSH is configured with.

Install

dsh plugin --profile web add @luisarg/memory-auto

Requires the vault server directory (memory-vault-server in this repository) and either uv

on PATH or Python ≥3.11: the bundled launcher.mjs runs the server with uv run and falls back

to a pip-managed venv (python3 -m venv + pip install -r requirements.txt) when uv is missing.

Skills

The package ships one skill, registered as a bundled provider (rank 600, the weakest in DSH's

local discovery table — a same-named skill in a user or project root still wins):

Command · Source · What it does

/checkpoint-auto · skills/checkpoint-auto/SKILL.md · Explains and steers the automatic capture: the git-commit, compaction, idle and session-end triggers, what an injected [memory-checkpoint] prompt means, and the provider / model / maxTokens / minTranscriptChars / enabled knobs

It deliberately does not restate the entry vocabulary: that lives in packages/memory-mcp's

checkpoint skill and in src/pure.ts, which builds the injected prompt so it keeps working with

no skill loaded. Registration happens before the enabled gate — the skill is how a session

learns that capture is off and how to turn it back on.

Configuration

Field · Purpose · Default

memoryPath · vault directory · $DSH_MEMORY_PATH or ~/.memories

serverDir · directory containing the vault server.py · $DSH_MEMORY_SERVER_DIR or $DSH_HOME/memory-vault-server

provider · LLM provider route (harness registry) · deepseek-official

model · LLM model id · deepseek-v4-flash

maxTokens · extraction output cap · 2048

minTranscriptChars · minimum transcript size to digest · 200

enabled · master switch · true

Environment variables used at load time: DSH_MEMORY_PATH,

DSH_MEMORY_SERVER_DIR.

Behavior

• session/created — registers the session and resolves the project name

(openspec/ present → directory basename; else package.json name; else

pyproject.toml [project].name; else first # heading of README.md;

else directory basename).

• session/disposed — digests the transcript, skipping sessions with no

activity. A dispose-time ctx.effect batch-digests any session still

pending, using the tracked activity summary as a stand-in transcript.

• agent/status with status idle — runs the auto-capture gate

(idleCheckpoint): skips sessions without activity or already delivered, and

on the first idle of a session also digests the tracked activity summary.

• session/event — tracks activity and queues checkpoints:
• tool/call (also accepted as tool_call) whose args.command matches

git commit queues a commit checkpoint.

• compaction/start queues a pre-compaction checkpoint (fires whenever there

is activity, even if one was already delivered).

• user/message and assistant/message are recorded as tracked activity.
• agent/pre-step — delivers the queued checkpoint by pushing it onto

payload.context (or payload.messages), and the agent writes entries with

the store_* MCP tools.

Extraction failures are retried with bounded backoff and logged; a failed

digest never takes the agent down.

Development

pnpm install
pnpm build
pnpm test

The bundle manifest (dsh.bundle) makes this package installable through the

standard dsh plugin add flow; the patch layer lives in cordis.patch.yml.