If you have used an AI Agent like OpenClaw in depth, you have probably lived through this "three-week trajectory":

• Week one (honeymoon): You pour the project's full context, decisions, and trade-offs into the Agent, and it can do anything for you
• Week two (unease): Every day you open it and spend 3-5 minutes reminding it what we are working on; deep judgments that span days are all gone
• Week three (downgrade): You stop asking deep questions and only use it to search and edit text — the "companion that thinks with you" has been demoted to a "lookup tool"

This is not a model capability problem; it is a memory system flaw. Hy-Memory exists to fix exactly this.

• Project: https://memory.hunyuan.tencent.com/
• Documentation: https://memory.hunyuan.tencent.com/openclaw/

What Hy-Memory Is

Hy-Memory is an OpenClaw memory plugin from Tencent. Its three-part core architecture lets an Agent, over long-term use, "remember at all, remember correctly, remember lightly, and understand you better".

On authoritative public benchmarks:

• Memory fragment count down 70%+, information density per memory up 45%+
• Token consumption for ultra-long-context processing down 35%
• Memory updates 20% faster

The Three-Part Core Architecture

Part one: the 6-layer memory framework

Hy-Memory does not stuff every memory into one table. It splits memory into 6 layers, one responsibility each:

Ask "what habits do I have when making big decisions" and the Agent checks the L5 mental models first; ask "where do I live" and a single L2 fact is enough. Each kind of question walks its own layer, so both retrieval efficiency and accuracy improve dramatically.

Part two: the System1 / System2 dual system

Borrowing from human-brain cognitive science, memory processing is split into two systems:

• System1 (day shift): Processes in real time as you send messages. Writes raw traces, extracts facts, updates the profile, compresses summaries — the fast L1-through-L4 operations
• System2 (night shift): Runs in the background at seconds-to-minutes scale. Extracts mental models, builds the knowledge network, predicts intent — the deep L5-to-L6 cognition

Why split? Because deep cognition is slow. Extracting one "decision mental model" can take 5-20 seconds. If every conversation waited 20 seconds, nobody could stand it. After the split, immediately usable memories are written by System1 in seconds, while deep understanding settles in gradually via System2 in the background.

Part three: evolution chains

This is Hy-Memory's most essential design, and also the easiest place to stumble.

Say you have discussed your fitness plan with the Agent for the better part of a year, and your stance took 4 turns: running -> HIIT (knee injury) -> pure strength (lost cardio) -> a mixed plan. Today you ask, "I want to add a new training style next month — what do you suggest?"

Three memory systems give three answers:

The overwrite school (keeps only the latest): Recommends CrossFit. But it does not know your knee was injured, and may well get you hurt again.

The hoarding school (keeps everything): Recommends HIIT. Because "running worked well" and "mixed training is stable" rank highest in vector similarity, and the "HIIT hurts knees" entry in between gets drowned out.

Hy-Memory evolution chains: Threads the 4 memories into one chain via supersedes pointers. When a search hits the newest entry, the whole chain unfolds automatically. The Agent gets the full path of how your stance evolved and will suggest steering clear of high-impact explosive training in favor of swimming or cycling.

Performance

On LongMemEval (500 questions across 6 capability dimensions) and PersonaMem (a realistic long-term conversation benchmark with 6000+ messages / 589 questions), Hy-Memory outperforms all comparable frameworks, leading on preference (+21.11pp), temporal reasoning (+9.63pp), and knowledge update (+21.37pp).

Write speed is in the same tier as mem0 and 8x that of Graphiti. It keeps only 1/3 the memory entries of mem0 and 1/4 to 1/5 those of Graphiti. Per-memory information density is 3-4x that of mem0.

Installation and Configuration

Install with one command:

npm config set registry https://mirrors.tencent.com/npm/ && \
openclaw plugins install @tencent/hy-agent-memory --dangerously-force-unsafe-install --force && \
openclaw hy-memory init

It uses Chroma as the default local embedded vector store, with data automatically persisted locally. No need to install Qdrant, no need for Docker — configure the API Keys for your LLM / Embedding and you are set.

Verify the installation:

openclaw hy-memory status

Three tiers to choose from

For a first install, going straight to Pro is recommended. Upgrading is just one switch — no re-integration needed.

Who It Is For

• OpenClaw power users fed up with "re-teaching the Agent every day"
• People who run project management or knowledge management on an Agent long-term
• Teams that need an Agent to genuinely understand their preferences and decision patterns