BOWLUNA/dsh-zcode-scribe

v1.0.0 · developed and verified against dsh >=0.1.5-rc.2 <0.2.0 || >=0.1.6-alpha.1 <0.2.0.

Long-term memory for DeepSeek Harness whose writer is a narrowed subagent.

Status: early, and read-only. The scribe_recall tool is registered, executes for real inside a live host, and a real model session on dsh 0.1.6-alpha.2 used it to answer from a memory room — docs/MEASUREMENTS.md carries the raw output behind every claim below. 101 tests pass. What does not exist yet: writes, extraction, and the narrowed writer, which is gated on milestone M0 — proving a subagent's capability set can actually be narrowed through a public seam. Nothing is injected into the prompt yet, so installing this changes what the model can ask for, not what it knows.

简体中文

---

The problem this is actually solving

The DSH plugin market lists 190 plugins under memory. Long-term memory is not an empty niche, and this project is not justified by a missing feature. It is justified by three missing safety properties, and it only takes the second slot in the table below because the first is already served by other people's work.

Property	Existing coverage
A local, bounded, inspectable store	Covered. engramory, dsh-memento, dsh-memoir and dozens more do this well. Install one of those.
Hard permission narrowing for the component that decides what to remember	None. No plugin removes capabilities from its writer; the closest analog, dsh-memento, gates writes behind human approval — a different and complementary mechanism.
Path safety on the write path (root-relative denylist, Unicode bidi/control stripping, NTFS alternate-data-stream truncation, containment)	None. Schemes that store in a database avoid it; every markdown-based option leaves it open.

This matters more every month. OWASP ASI06 — Memory & Context Poisoning joined the Agentic Top 10 for 2026, and the published numbers are not close:

• 98% injection success for a memory-poisoning payload (GhostWriter), ~60% activation even with polite phrasing, and 0% detection by existing one-turn injection filters.
• 95%+ (MINJA), 80%+ at under 0.1% poison rate (AgentPoison).

Anthropic's own memory-tool documentation says restricting operations to the memory directory "is not optional for any agent with write access to persistent storage", and advises scoping write permission to the sessions that actually add memories.

What makes the problem unavoidable is this: legitimate memory writing and malicious memory injection are the same operation. Same file, same write call — only intent differs, and intent is not observable.

So dsh-zcode-scribe does not try to detect intent. It removes capability, and constrains what remains:

the writer  →  cannot see:   shell · run_code · every MCP tool · network fetch/search
                            subagents · present · upload/attachment tools
             →  can do:      Read/Grep/Glob · Write/Edit only inside the memory room
                            rm only for a contained, absolute, non-glob, non-recursive .md

Two enforcement layers, because they fail differently:

Layer	Seam	Effect
Remove	ctx.tools.restrict({ deny }) on the writer's agent.ctx	the tool is absent from the writer's tool list — it cannot be talked into reaching for it
Constrain	ctx.tools.guard(exec => reason | undefined) on the writer's agent.ctx	argument-level policy for the tools that must remain, e.g. "read a file, but write only *.md under the memory root"

Both require a scoped context and both throw rather than silently degrade to global — which is what makes them usable as a security boundary instead of a convention.

What is borrowed

Almost everything. This project's contribution is the combination, not the parts:

• Claude Code — memory is an index, not storage; MEMORY.md capped at the first 200 lines or 25 KB; topic files loaded on demand, never at session start; an over-cap write succeeds and returns an error telling the model to rewrite the index rather than truncating silently; subagent memory in a separate directory.
• Anthropic memory tool — client-side storage ownership, /memories as a mapped prefix, mandatory path-traversal rejection.
• Anthropic managed memory stores — immutable version per change (audit + rollback), a hard capacity cap where writes fail loudly, and scoping write access.
• Memory-poisoning research (AM-Sentry, memory-risk scoring) — an admission policy before storage, trust tiers per entry, and holding a contradicting memory for a human decision instead of overwriting.
• The DSH ecosystem — engramory for the ctx.tools.guard() primitive, dsh-memento for loud-not-truncating budgets and session-frozen snapshots, dsh-memoir for prefix-cache-aware injection.

Full tables, sources and line-referenced seam evidence: docs/ARCHITECTURE.md.

Design in one picture

user turn ──▶ agent/post-step
                 │  skip if the main agent already wrote memory this turn
                 │  skip if the user turn carried no substantial prose
                 │  skip if the cursor did not advance
                 ▼  (coalescing: only the newest snapshot survives)
              mint the writer ──▶ restrict({deny}) + guard(...)   ← the whole point
                 ▼
              writer reads the manifest, writes candidates inside the room
                 ▼
              admission: path safety · cap (loud) · conflict hold · secret screen
                 ▼
              accept → versioned write → index rebuild → audit row
              hold   → needs a human decision

The main agent sees the index and a manifest of filenames plus descriptions. Topic bodies are never injected automatically — the model reads them on demand.

Install (when it is ready)

dsh plugin --profile web add dsh-zcode-scribe
dsh --profile web --dump-config | grep -A3 'id: scribe'    # rows collide ⇒ hard boot failure

Develop

node test/run.mjs          # all suites

Requires Node >= 20. No runtime dependencies — the plugin imports node: builtins and the host's peer packages only.

License

MIT.