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 0.7.0 (P3)

• Update-replace semantics: consolidation no longer coexists new and old versions of the same fact. A name-anchored matcher (prefix-tolerant token coverage, order-insensitive — ignore-err-x matches err-x-ignore) plus description-echo matching now routes matched candidates to an in-place UPDATE (reusing the existing name/scope and lifecycle) instead of skipping or duplicating. The consolidation prompt also instructs the model to reuse the existing name when updating.
• Matching core shared with the eval harness (src/matching.ts): the L3-validated mapping (name anchoring ∪ CJK-bigram coverage) is now the single implementation used by both product and runner — evaluation methodology feeding back into product, round three.
• New L3 metric — duplicates: counts gold facts still covered by >1 produced memory (un-replaced coexistence). After the change: mini 0 (was 1), standard products 26→25 with 2 remaining; precision 70→87.5% (mini). Recall on standard dipped to 75% in this single run — recorded honestly, run-to-run variance unmeasured.
• Panel UX pass: overview header (totals + pinned/muted + budget bar), colored type badges (user/feedback/project/reference), search filter, relative timestamps ("3 d ago"), icon actions (✏️📌🔇🗑), empty-state 3-step onboarding, per-group summary lines, current-workspace ✅ badge.

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:

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 o