dsh-journal

A DeepSeek Harness plugin that turns a plain JSONL file into a first-class journal tool with graded memory: entries carry a level, an expiry, and a review date, so a long-lived assistant can remember the right things for the right amount of time — and let the rest go on its own.

Why memory needs levels

A single undifferentiated memory file fails in two directions at once. Keep everything and the useful part drowns; keep only recent things and the relationship and the standing preferences — the parts that make an assistant feel like the same person — are exactly what gets dropped.

So every entry gets one of four levels, and the level decides its lifetime:

level	meaning	lifetime
S	relationship, emotion, shared experience, promises	never auto-deleted; pinned
A	durable facts, preferences, key decisions	365 days
B	project context, task progress	60 days
C	transient details, one-off operations	7 days

prune soft-deletes expired rows into archive/; nothing is ever physically removed by the tool. Pinned rows are exempt.

The useful question when writing is not "is this important" but "which of these is this, a skeleton or a file" — the rules for how someone talks to you belong in S, the port number you debugged yesterday belongs in B, and a handoff note for the next session belongs in A because it only needs to survive until you read it.

The tool

One tool, eight actions:

action	does
add	write one entry (level, title, text required)
list	list entries, optionally by level / status
search	substring search across title, text and tags
digest	the S + A summary — what a fresh session should read first
stats	counts per level, size, expired-but-present
review	push one entry's expiry out
prune	soft-delete everything past its expiry
render	rewrite JOURNAL.md as a readable digest

Typical use in an agent loop:

journal action=digest                          # at the start of every new conversation
journal action=add level=S title="..." text="..."
journal action=search query=备考
journal action=prune dryRun=true               # see what would expire, change nothing

Storage

One JSON object per line at <root>/journal.jsonl. Writes go through a temp file and a rename, so a crash mid-write cannot tear the store. A torn trailing line on read is skipped rather than failing the call.

Row shape:

{
  "id": "j-20261104-140902-479d",
  "ts": "2026-11-04T14:09:02+00:00",
  "level": "B",
  "kind": "fact",
  "title": "short one-line title",
  "text": "the body; facts, not reflections",
  "tags": ["..."],
  "pinned": false,
  "status": "active",
  "expires_at": "2027-01-03T14:09:02+00:00",
  "reviewed_at": null,
  "source": ""
}

The format is deliberately plain. Any script that can read JSONL can read this, and any script that can write one line can add to it.

Install

A DeepSeek Harness plugin only receives the host packages (@deepseek-ai/*) when it is installed as a link. An ordinary npm dependency fails to import and is silently skipped.

# 1) stable location, outside the profile
#    e.g. D:\DSH-Plugins\dsh-journal\0.1.0\

# 2) real junction (a cross-drive symlink can be created broken, silently)
New-Item -ItemType Junction `
  -Path   "$env:USERPROFILE\.dsh\profiles\desktop\node_modules\dsh-journal" `
  -Target "D:\DSH-Plugins\dsh-journal\0.1.0"

# 3) both lists in the profile's package.json must contain it
#    dependencies:        "dsh-journal": "link:D:/DSH-Plugins/dsh-journal/0.1.0"
#    dsh.profile.bundles: "dsh-journal"
#    Edit with Node, not a text replace: a stray comma makes the profile
#    unparseable and the harness will not start.

# 4) restart the harness

Configuration (in the bundle layer cordis.patch.yml):

key	default	meaning
root	D:/dsh-data/journal (Windows) / /dsh-data/journal	where journal.jsonl lives
toolName	journal	the tool's name in the model's toolset

JOURNAL_ROOT overrides root from the environment.

Test

node _selftest.mjs

Runs against a temporary store: add at each level, list, search, digest, stats, prune dry-run, render, review, and the two validation paths (bad level, empty title are both rejected).

Note on what it is not

This is not a vector store and does not do retrieval by similarity. It is a small, inspectable, hand-curated log — closer to a notebook than a database. The bet is that an assistant which writes down the right few hundred lines, and forgets the rest on schedule, beats one that hoards everything and searches it.

License

MIT