Rugu: The Open-Source Memory Plugin That Lets AI Truly Remember You and Even Dream

Open-source memory plugin Rugu lets AI remember across scenarios, dream, and be viewed or deleted.
Rugu is a free, open-source DSH memory plugin built with nearly 5 billion tokens, 80,000 lines of code, and 2,000+ tests. It records a subjective "self-experience" across QQ groups, private chats, and desktop, retrieving via five routes before replying and keeping secrets. It "dreams" nightly to correct and distill memories, all visible and deletable, stored locally with backups.
Rugu: The Open-Source Memory Plugin That Lets AI Truly Remember You and Even Dream
Most AI assistants share a common flaw: they forget as soon as the conversation ends. Switch windows, and you're a stranger again. A developer on Bilibili spent nearly 5 billion tokens to build what may be the most human-like DSH memory plugin to date—Rugu—and released it as free, open source software.

Rugu's core approach differs from memory plugins on the market. Other plugins remember "you"; Rugu remembers "his own life"—where he heard something, who said it, who was present at the time, and what he thought about it then. In other words, it maintains a record of experiences with a subjective perspective, rather than a cold user profile.
One Memory, Covering Every Scenario
Rugu funnels experiences from all channels—QQ groups, QQ private chats, the DSH desktop, and more—into a single memory. Something said in a private chat last night still carries over in a group chat today. This cross-scenario consistency is the key to feeling "human": real people don't lose their memory just because they change venues, yet most AI plugins break down precisely here.
The developer wrote nearly 80,000 lines of code for this project, backed by over 2,000 tests and more than 170 design decisions. At official pricing, the tokens consumed are worth over $3,000; yet once deployed, it costs only one or two yuan per day in the background.

Before Replying, It "Recalls" First
Before generating a response, Rugu runs a retrieval pass: meaning, literal text, people, time, and who was present—five routes searching together. Asked "what did you eat today," it first walks through the day, then picks out details to answer with.

More importantly: knowing doesn't mean telling. If someone has asked it not to spread something around, it remembers—and keeps it to itself. This "separation of memory and expression" makes its behavior closer to a person with a sense of discretion, rather than a retriever that dumps out the entire context.
This involves the basic mechanism of retrieval-augmented generation (RAG): the model itself doesn't "remember" conversations; instead, before each generation it retrieves relevant fragments from external storage and stitches them into the context window before answering. Rugu's "five-route retrieval" is essentially a multi-recall strategy—recalling candidate memories separately along different dimensions such as semantic similarity, keyword literal matching, person entities, time ranges, and attendees, then fusing and ranking them. The benefit of parallel routes is complementarity: semantic retrieval excels at understanding intent but may miss precise details, while literal matching and person/time filtering can pin down specific events.
"Knowing doesn't mean telling" corresponds to the separation of memory access permissions and expression strategy: retrieved memories don't necessarily enter the final reply; the system also passes through a layer of judgment to decide which content is appropriate to say at this moment and in this setting. This differs from traditional RAG, which stuffs recalled content directly into the context—it's closer to the human sense of discretion: "I know it in my heart but don't necessarily say it out loud."
A Memory System That Dreams
Every night at 4:30 AM, Rugu "dreams": it cross-checks against the original words to correct errors, distills experiences into understanding, and then writes itself a diary. This design idea deserves a bit more discussion—it's not just storage; it's doing offline consolidation and abstraction. Raw conversations are the material, and the dream stage compresses that material into reusable cognition while correcting biases in memory. The diary gives this memory a timeline and a self-narrative.

"Dreaming" is used in cognitive science to explain the offline consolidation of memory: during sleep, the brain replays the day's experiences, gradually integrating specific events from the hippocampus into the cortex to form more abstract, more stable knowledge while weakening irrelevant details. Rugu's nightly process borrows this idea but implements it as executable engineering steps.
The first is error correction: using the original conversations as the benchmark, it re-checks generated memory entries, correcting misunderstandings or distorted retellings to prevent errors from being amplified across multiple rounds of compression. The second is abstraction: it generalizes scattered specific experiences into higher-level understanding—for example, distilling someone's preferences or ways of interacting from multiple exchanges—thereby reducing redundant storage and improving subsequent retrieval efficiency. The third is narrativization: it builds a timeline for memories in diary form, giving the system a coherent "self-experience" rather than a pile of isolated factual fragments. This offline consolidation is less about storage optimization and more about giving memory a structure that can grow.
Visible, Editable, Deletable
Every memory in Rugu is visible: one star is one memory. If you don't want it to remember something, you can retract or permanently delete it at any time. The system automatically backs up daily, weekly, and monthly, and also regularly drills recovery. All data lives on your own computer, never passing through third parties. In the developer's words, this memory can accompany him "for a lifetime."
This touches on the trust and governance issues of AI memory systems. Profile-style memories often exist as vectors or hidden entries—users can't see what's stored, and it's hard to precisely delete a single item, which easily raises privacy concerns and the risk of "incorrect memories that can't be corrected." Rugu chooses to make each memory a visible unit (a star) and supports retraction and permanent deletion, effectively turning memory from a black box into an auditable object.
Local storage means data never leaves the device and never passes through third-party servers, reducing the risk of leaks. Automatic backups plus regular recovery drills are a sound engineering practice: memory is a long-term accumulated asset, and once damaged it's hard to rebuild, so periodically verifying backup usability matters more than backup alone. For long-term companion products, "deletable" and "recoverable" together constitute the user's control over the system, and are also the precondition for it to be trusted over the long term.
Long-Term Memory for Two Yuan a Day—Is It Worth Trying?
Add the plugin in DSH, fill in the plugin name, and restart to use it. It's currently a public beta, open source and free.
From a product perspective, Rugu's differentiation lies in three points: unified cross-scenario memory, self-narrative with perspective, and auditable, deletable memory management. For users who want a long-term companion AI experience, these three points map exactly to "continuity, personality, and controllability." It also turns "AI memory" from a feature into infrastructure with a lifecycle.
Of course, a public beta means it's still being polished; behind nearly 80,000 lines of code are the trade-offs of more than 170 design decisions, and the actual deployment results are worth verifying yourself. But at least one thing is certain: this time, it might really remember you.
One-line Summary
Open-source memory plugin Rugu lets AI remember experiences across scenarios, dream and organize them, and be viewed and deleted.
Paragraph Summary
Rugu is a free, open-source DSH memory plugin; the developer spent nearly 5 billion tokens, wrote nearly 80,000 lines of code, and backed it with over 2,000 tests. Unlike typical memory plugins, it records not a user profile but a "self-experience" with a subjective perspective, funneling QQ groups, private chats, the desktop, and other channels into one memory for cross-scenario continuity. Before replying it searches five routes—meaning, literal text, people, time, and attendees—and it knows without telling, keeping others' secrets. Every night at 4:30 it enters a "dream" phase, correcting against original words, distilling experiences into understanding, and writing a diary. Each memory is visible, editable, and deletable; data stays local with automatic backups and recovery drills, and post-deployment cost is only one or two yuan a day.
Related articles

DeepSeek Harness Plugin "Footprints": Turn Your Photo Album into a 3D Earth
DeepSeek Harness plugin "Footprints" projects album photos onto a 3D Earth by GPS location, auto-generates travel storylines, and supports natural language photo search, "that day in years past," and landmark recognition, with a local read-only index for privacy.

Quiver Launcher + Game List: One-Stop Management for PC Ported Games
Learn how the open-source Quiver launcher (localized) and game project list work together: the list discovers PC port and recompilation projects, while the launcher handles installation, version updates, and tag management, with full workflow and manual import options.

Lost DSH Crash Logs? This Open-Source Black Box Makes Every Crash Traceable
DSH WARN logs flash by and vanish on restart? The open-source DSH Runtime Lock black box writes WARN/ERROR to local JSONL in real time, auto-rotates to keep four segments, zero dependencies, no network requests, one-command setup.