AI Learns Like Humans: Open-Sourcing "Cerememory," a Living Memory Database Based on Brain Mechanisms

AI Learns Like Humans: Open-Sourcing "Cerememory," a Living Memory Database Based on Brain Mechanisms

Co-R-E Inc. (Headquarters: Shinjuku-ku, Tokyo; Representative Director: Masato Okuwaki), an AI connective company, has released "Cerememory," a memory layer that implements brain structures to provide AI agents with a "living memory" database. It is open source and available for anyone to use.

Official Website: https://co-r-e.github.io/cerememory-docs/ja GitHub: https://github.com/co-r-e/cerememory

### Why Reference the Human Memory System?

Currently, developers worldwide are grappling with the challenge of giving AI agents long-term memory. Various methods are emerging, including proprietary memory functions, RAG using vector databases, and context engineering. Many excellent implementations are already available.

However, looking at these efforts from a broader perspective, we arrived at one question: "When considering the ideal state of AI memory, what would be the ultimate goal?"

The answer to this question was surprisingly close: a living memory system, just like ours.

Human memory is not merely a mechanism for storing and retrieving information. Important experiences remain vivid, trivial events fade, similar memories blend, and recalling one memory can trigger a chain of related memories. Everyone has experienced childhood memories resurfacing due to a nostalgic scent.

During sleep, the day's events are organized and integrated into long-term knowledge. It is an extremely advanced information processing system, refined over hundreds of thousands of years of evolution.

The question Cerememory addresses is not "how to give AI memory," but rather "to what extent can we incorporate the excellent memory mechanisms that humanity has acquired over long periods into the AI's memory layer?" Cerememory's approach is to implement phenomena observed in neuroscience research as an execution system, not merely as data structures.

### Three Design Principles of Cerememory

- **Treat memory not as "static storage" but as "dynamically alive"** Human memory is not stored statically like a hard disk. Important things are strengthened, unused things fade, and memories subtly change each time they are recalled. Cerememory inherits this characteristic, implementing dynamic processes such as decay, interference, reactivation, the reorganization of memories during dreams while sleeping, and integration as an execution system, not merely as data structures. Memory is not fixed at the moment it is written but continues to move dynamically over time.

- **Memory should have "reasons" as well as "content"** Many people have likely struggled to understand the intentions behind an AI agent's actions. When a person remembers something, that memory is accompanied by a meta-context of "why it was considered important" or "how that decision was made." Cerememory incorporates a meta-memory plane into every AI agent's memory record, structuring and recording "why it exists." This includes intentions, grounds, evidence, alternatives, and decisions. Memory can be traced not only by the AI agent's actions but also by their reasons.

- **The memory layer should be independent of the AI** Human memory belongs to the person themselves and is not subservient to a specific conversational partner. Cerememory adopts the same stance. Through its unique protocol "CMP*," it can access the same memory layer from any LLM, such as Claude, GPT, or Gemini. Data is stored locally and can be exported at any time. Even if you switch AIs or cancel a service, your memories remain with you. The memory layer should not be owned by a specific AI or vendor.

*Note: CMP and MCP are different protocols. While their names are similar and often confused, CMP (Cerememory Protocol) is a proprietary protocol defined by Cerememory for reading and writing memories, whereas MCP (Model Context Protocol) is an open specification proposed by Anthropic for connecting LLMs with external tools. They are not competing but rather have a relationship where "MCP carries CMP" within Cerememory. The structure is such that MCP clients like Claude Code or Cursor deliver CMP messages internally to the Cerememory engine.

### "5-Store Architecture" Directly Incorporating Brain Structure into Design

The human brain processes different types of memories in separate areas.