AI News

Red Bear AI's open-source memory project MemoryBear surpasses 10,000 GitHub stars

China.com TechnologyUpdated Nov 10 14:18

AI summary

Red Bear AI's open-source project MemoryBear has surpassed 10,000 stars on GitHub. The project adopts a layered architecture of working memory, short-term memory, and long-term memory; long-term memories are stored as a spatiotemporal semantic graph with timestamps, sources, and versions, and support weight-based decay and conflict validation. Red Bear AI has also released the core repository, personal client integration methods, MCP protocol adaptation, and a self-developed evaluation benchmark, and enterprises can deploy it privately.

Why read this

AI score 55

The article outlines the design of memory layering, forgetting decay, and source traceability, helping readers judge the difference between a memory foundation and retrieval-based approaches.

Topics

Tags

Body · AI translation

Red Bear AI's open-source project MemoryBear has surpassed ten thousand stars on GitHub. What this number shows matters more than its popularity: developers are treating AI memory as a foundational capability that needs to be built separately, rather than merely looking at whether it is an incidental feature of a model.

Over the past two years, the parameter counts, context lengths, and inference speeds of large models have climbed steadily, but memory has not kept pace. Most products lose context when a session ends, and when users ask again a few days later, they have to explain everything from scratch. Red Bear AI fills this gap with a self-developed memory engine, and has chosen to open-source its core capabilities.

MemoryBear surpasses ten thousand stars on GitHub

Behind ten thousand stars, what developers lack is governable memory

The hard part of AI memory is not storage, but management. There is no shortage of memory-related projects in the open-source community, and most of them solve the same thing: retrieving previous conversations. This approach can handle demos, but it cannot sustain long-term use. The real difficulty is how to manage memory once it comes in: when it should be updated, what should be forgotten, whether a conclusion can be traced back to its original source, and how to isolate it when multiple people share it.

MemoryBear builds these matters into its architecture and has also put its engineering implementation into a public repository. Stars are votes cast by users, and what they are voting on is whether this thing can be used directly, not whether the concept sounds new.

Memory has long been made into a layer of retrieval

Equating memory with retrieval is a very natural choice in the industry. Conversation caching handles the current context, vector retrieval handles recall of similar fragments, and external documents handle knowledge supplementation. Together, these three can cover a considerable number of scenarios.

The gap lies in continuity. Information has an order, versions, and expiration, and the fragments retrieved lack a time coordinate, so contradictory old and new content will remain in the database at the same time. When users change a plan or switch preferences, the system has no mechanism to determine which version still counts, and answers will waver between old and new.

In actual work, the problem is even more obvious. A project spans several quarters, requirements have changed several rounds, early conclusions are already obsolete, and old documents remain on the shared drive; when the contact person changes, the successor cannot tell which judgment is still in effect. Retrieval can find all five documents together, but it cannot say which one still counts.

MemoryBear takes a layered approach. Working memory corresponds to a single task and manages what is needed for the current step, conveniently filtering out pleasantries and redundant phrasing and keeping only elements such as entities, relationships, time, constraints, and confidence. Short-term memory takes over the filtered content, performing merging, completion, and conflict validation in the background, while parts that have not been awakened for a long time decay by weight. Long-term memory is stored as a spatiotemporal semantic graph, with each record carrying a timestamp, source, and version, and retrieval follows three paths: temporal, causal, and relational.

Further subdivided, this system can also be split into five layers: perceptual, working, explicit, implicit, and emotional, for discussing finer mechanisms. When explaining it externally, a rough three-layer division is enough: first filter, then consolidate, and finally form a graph. Each layer has its own writing and elimination rules, and only then can memory be governed.

After layering, forgetting becomes something that can be configured. Memories without new supporting evidence for a long time are downweighted, and memories that are frequently reused and repeatedly verified are consolidated. Decay does not equal deletion; they can still be retrieved when needed. Each record carries a confidence level, with facts, inferences, and uncertain information labeled separately, so users can see which parts are speculation.

Red Bear AI has released MemoryBear's core repository, the access method for personal clients, MCP protocol adaptation, and a self-developed evaluation benchmark through public channels. Enterprises can deploy it privately, developers can integrate it directly, and the evaluation data and methods are made public, so results can be reproduced and verified.

During multi-agent collaboration, memory is divided by task, and each agent receives only the part it needs, so team sharing efficiency and data boundaries can both hold.

The value of this matter is not whether the code is public, but the barrier to use. If memory were rebuilt separately by each company, the industry would repeatedly invest at the same layer; with a publicly available foundation, small teams can also make their products remember users.

The standard for open-sourcing something is not that the code has been released, but that others begin to use it to build their own products. Ten thousand stars is the beginning of this memory system being put to use.

Source

China.com Technology