How Tencent's Hy-Memory Plugin Gives AI Agents Long-Term Recall

The layered approach gives each type of memory a dedicated space, from the rawest interaction data to high-level mental models .

This segmentation is meant to solve the problem of "searching a haystack for a needle." By not mixing raw logs with personality profiles and abstract intentions, retrieval can be more targeted and contextually appropriate.

System 1 / System 2 Dual Processing

Hy-Memory's processing model borrows directly from cognitive science, splitting tasks between fast, real-time operations and slower, more analytical background processes .

• System 1 ("Day Shift"): Triggered the moment a user submits input, System 1 handles L1 through L4 in real time. It writes raw traces, extracts atomic facts, updates identity profiles, and compresses session summaries—all without slowing down the primary chat loop .
• System 2 ("Night Shift"): Operating asynchronously on a scale of seconds to minutes, System 2 runs in the background to manage the heavier cognitive tasks of L5 and L6. This includes extracting mental models, constructing knowledge networks, and predicting user intent .

This separation is key to balancing speed with depth. A user doesn't wait for deep analysis during a conversation; they only feel the lightweight impact of L1-L4, while the agent continues to "think" and refine its understanding behind the scenes.

The Evolutionary Chain Mechanism

Perhaps the most novel piece of the architecture is how Hy-Memory handles updates and corrections. Instead of simply overwriting or deleting old information, new memory entries use a supersedes pointer to link back to the data they are replacing or refining .

This creates an evolutionary chain of related facts. When a search hits any node in the chain, the entire context can be unfolded. The system places the most recent, correct judgment at the head of the chain, but older, superseded versions are not discarded—they are simply pushed back, available for full context if needed. The result is a memory system designed to evolve alongside a long-term user relationship without the fragmentation or data loss typical of simple append-only logs .

Measured Performance: Less Noise, Higher Signal

The architectural claims are backed by performance numbers from public benchmark testing reported at launch. The central theme is memory densification: packing more useful information into fewer records and tokens .

• Memory Fragmentation: Total memory entries were reduced by more than 70%, directly tackling the problem of cluttered, fragmented memory stores .
• Information Density: The information carried by each stored memory entry increased by over 45%, making each recall operation more effective .
• Token Efficiency: Token consumption for processing long-context information dropped by 35%. This is a critical cost and context-window saving .
• Write Speed: Hy-Memory's write speed was clocked at 8× faster than Graphiti and comparable to Mem0, meaning the memory system does not become a bottleneck in live conversations .
• Update Speed: The speed of updating existing memories improved by 20% .

In practical terms, this means an OpenClaw agent running Hy-Memory should feel snappier, recall more relevant information, and consume fewer resources, with a memory count that could be as little as 1/3 of Mem0's or 1/4 of Graphiti's for the same user .

How It Differs From OpenClaw's Default Memory

By default, OpenClaw agents write everything to Markdown files: a MEMORY.md for long-term facts and dated files in a memory/ directory for daily session logs . This two-layer system is functional but limited—it requires manual curation for long-term accuracy and can easily become noisy .

Hy-Memory replaces this flat-file model with a structured, automated cognitive pipeline. The shift from a passive file-based approach to an active, layered, and evolutionary system is what Tencent means by a "second brain." It’s not just storing data; it’s curating, compressing, and contextualizing it over time.

Deployment and Configuration Reality

Hy-Memory is distributed as an OpenClaw plugin available through the marketplace, and it runs a local Python process to service memory read/write operations . According to Tencent's official documentation, it can run purely locally with SQLite, or it can optionally be connected to Tencent's cloud vector database (TCVDB) as a remote backend .

A note on public claims: while the term "three-tier" is prominent in marketing materials, the available launch reports describe the core design as being "centered on evolutionary chains" rather than clearly delineated Lite/Pro/Ultra deployment tiers. Details on any such packaging, Docker dependencies, or specific default configurations like Chroma are not substantiated by the sources provided in the launch documentation . The verified headline is that Hy-Memory is a specialized plugin for the OpenClaw ecosystem, focused on solving a concrete technical problem through architectural rigor.