curated collections / cross-session-memory
DSH Preference Memory: Plugins That Remember You
Eleven maintained plugins, ★13 to ★996, that remember you between sessions — your preferences, your principles, your project context. Stored: the person, not documents.
A knowledge base remembers documents; RAG retrieves passages. This cluster remembers you. DeepSeek Harness sessions start from zero by default — your preferences, your red lines and the architecture you settled on last Tuesday all vanish unless a plugin carries them over. Right now eleven maintained plugins do exactly that, from mindmemos (★996), which auto-recalls your user and task context before every turn, down to dsh-mirror (★13), which keeps a bounded, explainable profile of your principles. They split into three builds — zero-prompt recall engines, knowledge-graph cores and governed preference profiles — compared below.
New to memory plugins? Install one auto-recall plugin, work a normal day, then ask it what it knows about you. If the answer is right, keep it; if it is wrong, fix the memory file — that one edit is the entire maintenance routine.
68 plugins3,784 stars
Synced from our GitHub list · Catalog updated 2026-09-27
Why it matters now
The thickest supply on the catalog: eleven maintained plugins between ★13 and ★996, every one installable into a profile today. Memory is the most-starred theme in the dsh ecosystem, and preference memory is where the new entries land.
It attacks a cost you pay daily: re-explaining yourself. Preferences, conventions and task state retyped every session are pure token waste and pure friction — a preference plugin pays for itself within a week.
Clear boundaries: this is not a knowledge base (that stores documents) and not context compaction (that shrinks the current session). Cross-session memory stores who you are and what you decided — the two things no other layer keeps.
In this collection
Editor’s picks lead, then the rest of the theme is auto-discovered by GitHub topics and ranked by stars. Catalog data, refreshed on every sync.
- 1mindmemosby mindscale-noah
Gives dsh agents persistent cross-session memory: auto-recalls relevant user/task context before each turn and writes back lessons after each turn, evolving memory via schema learning and skill distillation
Editor’s pickCuratedSessions & Messages99699 - 2graph-memoryby adoresever
Knowledge-graph memory core native to DSH (Cordis adapter, cross-session triples).
Editor’s pickCuratedKnowledge & Research62895 - 3mnemonby mnemon-dev
LLM-supervised persistent memory for AI agents — graph-based recall, cross-session knowledge, single binary. Works with DeepSeek Harness, Claude Code, OpenClaw, and any agent runtime.
Editor’s pickCLI & Terminal59672 - 4dsh-memoryby furongjun-1999
Cross-session long-term memory infrastructure for DSH agents: memory write/retrieve, continuous learning, auditable trust ('智能论 v3.2').
Editor’s pickCuratedSessions & Messages25620 - 5dsh-noemaby zseven-w
Noema long-term memory plugin for DSH: durable, inspectable agent memory with recall tools and a settings page.
Editor’s pickCuratedSessions & Messages1289 - 6dsh-meow-memoryby phant0meow
Cross-session seven-layer memory (soul/user/project/fact/lesson/topic/rules) with BM25 retrieval and per-window dream consolidation for DSH agents
Editor’s pickCuratedSessions & Messages12411 - 7mnemeby slow-stack
🧠 The memory that dreams — cross-session memory for DeepSeek Harness. Offline & private, auto-consolidates in its sleep (autoDream), visualized in a memory panel.
Editor’s pickCuratedKnowledge & Research12319 - 8dsh-mementoby perrylink
Bounded, layered, approval-gated, auditable cross-session memory for DeepSeek Harness (capability seam: ctx.memory + SQLite provider + memory tool + frozen snapshot injection)
Editor’s pickCuratedKnowledge & Research1203 - 9dsh-project-brainby yj-liuzepeng
Persistent project intelligence and memory plugin for DSH: architecture analysis, cross-session context, TODOs, and optional hybrid retrieval
Editor’s pickCuratedKnowledge & Research890 - 10dsh-auto-memoryby aik358
Three-layer auto-memory companion for DSH web GUI (calendar, greetings).
Editor’s pickCuratedKnowledge & Research786 - 11dsh-mirrorby webkubor
Persist bounded, explainable user preferences and inject them into future system prompts with a Memory tab.
Editor’s pickCuratedKnowledge & Research138 - 12dsh-mnemonby omdsh-dev
Composable, view-based memory for DeepSeek Harness. Pluggable sources and strategies, with three-tier memory out of the box.
CuratedSessions & Messages41733 - 13stratagate-agentmemoryby diqierjia
DSH 跨会话记忆插件：本地优先（SQLite）自动捕获对话与工具结果、L0-L5 分层存储、事件/元素卡片+知识图谱+时间线 UI、BM25+结构化混合检索、证据门控防无据作答、来源追溯
CuratedSessions & Messages977 - 14dsh-plugin-memoryby justhalfbit
Stores project facts, decisions, lessons, and preferences across sessions with searchable tools and on-demand topic files.
CuratedKnowledge & Research131
Auto-recall: how dsh cross-session memory usually works
These engines watch every turn and re-inject the relevant slice of your context before the model speaks — no slash commands, no manual recall. The most common build, and the easiest to live with.
mindmemos (★996)
The heavyweight: auto-recalls relevant user and task context before each turn and writes lessons back after it, so the profile evolves while you work. Memory quality improves through schema learning and skill distillation rather than growing into an undifferentiated blob.
Trade-off: it decides what to remember. That is the point — but review the store after your first week.
dsh-auto-memory (★78)
Proactive associative recall injected before the model speaks, with three-layer auto-consolidation and skill crystallization on top. Built around the DSH web GUI experience — calendar and greetings included — so it feels less like infrastructure and more like a companion.
The lightest of the auto-recall three and the easiest to try first.
dsh-noema (★128)
Durable, inspectable memory: recall tools the model can call, plus a settings page where you read what was actually stored. Noema’s bet is that memory you cannot audit is memory you should not trust.
Best fit when you want cross-session recall and a clear view of the store in the same plugin.
dsh-memory (★256)
furongjun-1999’s cross-session memory infrastructure: write and retrieve tools, continuous learning, and an unusually explicit architecture — the repo frames itself as a white-box AGI experiment with metacognition loops and auditable guardrails.
The most experimental of the group; the memory primitives themselves are solid.
Knowledge-graph memory: structure instead of vibes
These store memory as typed triples and graph nodes rather than loose notes, trading a little setup for precision: recall an entity and you get its relationships, not a paragraph that happens to mention it.
graph-memory (★628)
A knowledge-graph memory core native to DSH — the Cordis adapter persists cross-session triples that any dsh agent can query. Presented at a Tsinghua University seminar in April 2026, rare academic visibility for a plugin at this layer.
Choose it when you want memory to answer “how does X relate to Y”, not just “what did we say about X”.
mnemon (★596)
LLM-supervised persistent memory with graph-based recall, shipped as a single binary and working across DeepSeek Harness, Claude Code and other agents — so one memory can follow you between tools.
The supervision layer is the differentiator: an LLM reviews what enters the graph instead of trusting every write.
mneme (★124)
“The memory that dreams” — offline and private, it auto-consolidates in its sleep (autoDream): between sessions it merges, prunes and rewrites its store, then visualizes the result as a mind map.
Pick it if you want a memory that stays tidy without manual curation.
Preference memory: bounded profiles of you
The strictest build: memory is written into named, bounded layers — preferences, principles, red lines — with approval gates, capacity caps and human-readable files, so “what does it know about me” always has an answer.
dsh-meow-memory (★124)
Seven-layer SQLite store — soul, user, project, fact, lesson, topic, rules — with BM25 retrieval and per-window dreaming to consolidate whatever accumulated.
The layered schema is the feature: each kind of memory lives where it belongs, and every layer reads as plain entries you can edit.
dsh-memento (★120)
Bounded, layered, approval-gated, auditable. Memory sits behind a capability seam (ctx.memory plus a SQLite provider), nothing is written without approval, and every entry can be traced after the fact.
The pick for shared machines, and for anyone whose threat model starts with “what exactly did it write down about me”.
dsh-mirror (★13)
The purest preference-memory plugin: the model notes your principles, red lines and working style, then reuses them across sessions — with a capacity cap, deliberate forgetting, and a stated reason attached to every entry.
Small star count, sharp idea; the closest match in the catalog for the literal search “dsh 偏好记忆”.
dsh-project-brain (★89)
Remembers the project rather than the person: architecture analysis, cross-session context, TODOs and optional hybrid retrieval, so the brain attaches to the repo you are working in.
Pair it with a person-facing profile plugin if you want both — the two do not overlap.
Frequently asked questions
How is a preference-memory plugin different from a knowledge base or RAG plugin?
Direction of storage. A knowledge base or RAG plugin stores reference material you point it at and retrieves passages on demand. A preference-memory plugin stores observations about you — decisions, conventions, working style, red lines — and re-injects them so every new session starts already knowing you. The two coexist fine: the knowledge base answers what the docs say, the memory layer answers what you prefer.
Does a cross-session memory plugin leak my private data?
The plugins in this collection store locally — a SQLite database or Markdown files inside your DSH profile directory — and the memory layer itself uploads nothing. Before installing, check three things: whether writes are gated (dsh-memento approves every write), whether you can inspect and edit the store (meow-memory and noema expose readable layers), and whether any companion service is involved. The memory is only as private as your machine, which is exactly the point of local storage.
Can I run two memory plugins at the same time?
Pick one primary memory layer. Two auto-recall plugins both inject context before each turn, so you get duplicate injections, contradictory summaries and doubled token spend — with no way to tell which memory answered. A memory plugin pairs fine with non-overlapping layers such as context compaction, and a graph core plus a preference profile can be combined deliberately, but two plugins of the same build will fight.
How do I uninstall a DSH memory plugin?
Uninstall the plugin like any other — DeepSeek Harness’s plugin manager handles removal. The memory it wrote is separate: SQLite files or Markdown inside your profile directory, and they survive the uninstall on purpose. Delete that data too if you want a clean break, or keep it — a successor plugin can often read it. Step-by-step: the uninstall guide linked below.
Which memory plugin should a beginner start with?
mindmemos if you want the most automatic path — it recalls and writes back with no configuration. dsh-auto-memory if you mostly use the web GUI. meow-memory if you want to read what it remembers (seven named layers). dsh-memento if privacy comes first and you want every write approved. Start with one, work a day, then inspect what it actually stored before adding a second layer.
Explore related categories
Explore the full directory
Hundreds of plugins across every category — or submit your own and get listed.