dsh-auto-memory

English | 中文

Your dsh agent forgets everything you tell it. Every. Single. Session.

Fix it with one command. Claude Code-style persistent memory for DeepSeek Harness — native, zero servers, zero embeddings, zero setup.

dsh plugin --profile demo add dsh-auto-memory

Say "Remember: I'm a Python backend engineer preparing for interviews" today — open a brand-new session tomorrow, ask "what do you know about me?", and it remembers.

---

What's new in 0.6.0 (P2)

• Harness-sessions evaluation datasets (evals/datasets/): synthetic multi-session developer transcripts with gold-label facts — durable / ephemeral-trap / updated labels, evidence sessions, abstention probes. The coding-agent memory scenario no public benchmark covers.
• L1 injection recall (zero-API runner): 100% on both datasets at the perfect-consolidation ceiling — the injection side loses nothing.
• L3 consolidation quality (one LLM call per session, same code path as the plugin): mini recall 100% / update-correctness 100% / pollution 0%; standard recall 83.3% / precision 84.6% / update-correctness 100% / pollution 0% (deepseek-chat, temperature 0). Ephemeral traps were correctly refused in every run.
• Product fix surfaced by the eval: consolidation now writes title/description in the transcript's language (Chinese transcripts used to produce English memories — measured, then fixed).
• Community: soft-mute (third state between keep and delete, #10), panel current-workspace fix (#9), CJK bigram echo-dedup (#5), editor busy/error UX (#8), GB11643 checksum before id-card redaction (#7).

What's new in 0.5.0 (P2)

• Three-factor ranking (Generative Agents formula, embedding-free): injection index and memory_list now order by pin → importance × recency (0.995^days, day-granular so the index text stays stable within a day for KV-prefix caches). Auto-consolidation has the model assign importance 1-10; types carry sensible defaults.
• Security pack: every write path now passes through redaction ([REDACTED:github-token/npm-token/openai-key/aws-key/phone/id-card] — the exact shapes that leaked in this project's own history) and protocol-tag escaping — the injected section is wrapped in <memory_context> and content can no longer forge the wrapper. Consolidation dedupes near-identical descriptions (Jaccard ≥ 0.7) instead of piling up echoes.
• Eval arc completes: probe retention under half-budget pressure is now 38% (positional) → 80% (pinned) → ≥90% (three-factor, unpinned) — same budget, same seed, better memories.

What's new in 0.4.0 (P2)

• Web UI memory panel (the screenshot above): a dual-half plugin — the host registers authenticated /api/auto-memory/* routes over the existing locked store, the browser half registers a sidebar entry and a management page (grouped list with types/pins/read counts, per-scope capacity bars with budget warnings, inline editing, pin toggle, delete-with-confirm). Works in Web and desktop clients from one bundle; non-Web compositions stay no-op.
• Panel bugfixes: routes now register via a connection-ready callback (host load order used to skip them), and the editor uses official theme variables (--dsw-alias-*) instead of a hardcoded white — readable in dark mode.
• Package hygiene: the browser half ships as lib/client.js (17 kB) behind a dsh.client manifest; total install stays ~32 kB with zero runtime deps.

What's new in 0.3.0 (P2)

• Pinned memories (pinned: true on memory_write): pinned entries lead the index, survive budget truncation, and are exempt from staleness eviction — a trust anchor the user controls.
• Eval-driven fix: the injection budget now covers the whole section (index + guidance); it used to overshoot by ~800 bytes. Caught by the new deterministic evaluation layer on its first run.
• Deterministic eval layer (evals/) in CI: injection budget curves, eviction zero-misfire, link-expansion bounds, and a signal-to-noise characterization — which pinned-priority truncation then improved from 38% → ≥80% probe retention under half-budget pressure. Same budget, better memories.

What's new in 0.2.0 (P1)

• Auto-consolidation (autoSummarize: true): when a root session ends, a background LLM pass extracts durable new facts from the session and files them as memories — deduplicated, capped, fully silent on failure. Claude Code doesn't do this automatically.
• Forgetting & eviction: every memory carries lifecycle metadata (created/updated/reads); memory_read counts references; staleAfterDays soft-hides zero-reference stale memories from the injected index (files kept); memory_prune lists (dry-run) or deletes aged memories.
• Recall expansion: memory_read resolves [[name]] cross-links one level and attaches linked summaries.
• memory_delete_all — guarded by tools/pre-execute human approval: the model cannot self-confirm irreversible bulk deletes.
• Hardened by a second adversarial review (11 agents): single-lock clear (no concurrent-write escape), conditional index rebuild on touch (no O(N) amplification), session-start stale refresh, subagent capture cleanup, abortable consolidation.

Tools: memory_write / memory_read / memory_list / memory_delete / memory_prune / memory_delete_all.

Claude Code has this. dsh didn't. Now it does.

DeepSeek Harness is the hottest open agent harness on GitHub right now — models, tools, sandboxes, everything is a plugin. But it ships with no memory subsystem at all. The official answer is three default-off MCP configs to third-party servers, which the official docs themselves qualify: not auto-injected, no forgetting policy, substring-only search. Your agent has amnesia by design.

dsh-auto-memory closes that gap natively:

	MCP bridge approach	dsh-auto-memory
Memories injected into every system prompt, automatically	✗	✓ (zero tokens when empty)
Typed memories: user / feedback / project / reference	✗	✓
Workspace + user scope layers — no cross-project leakage	✗	✓
Crash & concurrency safety (cross-process locks, orphan recovery)	—	✓
External services / databases / embeddings required	✓✓✓	none — just plain Markdown files

Memories are ordinary files under $DSH_HOME/memory/ — hand-editable, grep-able, git-friendly, yours.

One minute to feel it

node scripts/demo.mjs   # no API key, no browser: watch write → index → inject → recall → forget

Or for real, in a chat: tell your agent things worth remembering. The model calls memory_write / memory_read / memory_list / memory_delete, following Claude Code's write discipline: dedupe-and-update over piling up, absolute dates only, [[name]] cross-links, feedback memories carry Why: / How to apply: lines.

What the model actually sees

Every request, one system-prompt section (order 4000) carries the index — re-evaluated per step, byte-budgeted, and gone entirely when the store is empty:

# Persistent memory index
## Project memories
- [压测过 PostgreSQL](id-generator-benchmark.md) — psycopg2 连接池有踩坑经验 (2026-09)
- [用户是 Python 后端工程师](user-prefers-python.md) — 正在准备面试; 偏好中文交流

Chinese titles, YAML frontmatter, one file per memory — exactly the Claude Code MEMORY.md model, rebuilt natively on dsh's prompt-assembly pipeline.

Hardened before first release

This plugin survived a 12-agent adversarial code review (680k tokens of source-level scrutiny) before v0.1.0. Five production-grade traps were caught and fixed — with regression tests — including two that would have been field incidents:

• The NTFS silent destroyer: a memory named memory collides with MEMORY.md on case-insensitive filesystems — the write succeeds while destroying the record. Blocked by a reserved-name guard.
• The poisoned-prompt bomb: three literal {{{ }}} braces in any memory could crash every model request in the workspace — with no way for the model to self-recover. Neutralized by a converging sanitizer.

Plus: orphaned-lock self-healing (Ctrl+C can't brick your memory store), symlink-read protection, malformed-file tolerance, stable index ordering to protect KV-prefix caches, and a strict no-custom-session-events policy (they make dsh sessions refuse to resume).

Measured, not just claimed

A deterministic evaluation layer (evals/) runs in CI — no LLM, fully reproducible:

• Injection budget holds at any scale: 20/50/100/200 memories → the injected section stays ≤ 4 KB (4065/4048/4018/3940 bytes measured), with truncation markers; empty store injects 0 bytes.
• Eviction never misfires: four-class mixed scenario — only stale-zero-read memories get hidden; zero files lost; one read revives.
• Known limitation, pinned as baseline: budget truncation is currently positional (index order), not relevance-ranked — probe retention under half-budget pressure drops to ~38%→10% as N grows. Pinning fixes it for what matters: pinned probes retain ≥80% at the same budget (0.3.0); full relevance ranking remains on the roadmap.

This evaluation layer already caught a real bug: the byte budget used to exclude the policy text, overshooting by ~800 bytes (fixed, regression-tested).

78 tests (incl. a deterministic eval layer). 0 runtime deps beyond yaml. 15 kB installed.

Install

dsh plugin --profile demo add dsh-auto-memory   # from npm (prebuilt)
dsh --profile demo                               # restart the profile

From source: npm install && npm run build && dsh plugin --profile demo add /abs/path. Requires @deepseek-ai/dsh >= 0.1.5-rc.2 (Node ^22.19 || >=24).

Configuration

Override in your profile's cordis.patch.yml (config replaces wholesale):

- id: auto-memory
  config:
    maxBytes: 4096          # injection budget (index + guidance + wrapper)
    memoryDir: D:/memories  # default: $DSH_HOME/memory
    enableUserScope: true   # false: user layer off on every path
    autoSummarize: false         # true: on root-session end, a background LLM files durable new facts
    autoSummarizeMaxMemories: 5  # cap per consolidation
    autoSummarizeMaxTokens: 2048 # consolidation output cap
    staleAfterDays: 0            # >0: soft-hide zero-read memories older than N days (files kept)

How it works (60 seconds)

• Write: tool execute → name normalized to [a-z0-9-] (reserved names rejected) → cross-process file lock (official dsh-atomic-write) → atomic write → full index rebuild inside the lock.
• Inject: one dynamic section re-evaluated on every step assembly; reads the index synchronously, enforces the byte budget, neutralizes {{. Tool writes take effect on the very next request — no restart, ever.
• Audit: no custom session events (third-party types make dsh refuse to resume); everything flows through standard tool/call / tool/result.

Deep dives: design decisions · postmortem: shipping a PR to awesome-dsh-plugin

Roadmap

• P0 — typed store, four tools, prompt injection, scoped layers, crash safety
• P1 — auto-consolidation on session end, forgetting & eviction, recall expansion, human-gated bulk delete
• P2 — Web UI memory cards, token-cost / recall-quality benchmarks

License

MIT