Hy Memory

Hy Memory (written Hy-Memory in repository and model names) is a long-term memory system for AI agents developed by Tencent's Hunyuan team (branded "Tencent Hy"). Its second generation, Hy Memory 2.0, was released as open source in mid-October 2026 as three separately usable layers: a Python memory engine and SDK (hy-memory), a specialized memory-writing model called Hy-Memory-30B-A3B, and MemHub, a local-first desktop and command-line app that pools conversation history from several AI coding and agent tools into one shared memory.^[1][2][6][7] The SDK, the model weights and MemHub are published under the Apache License 2.0.^{[1][2][11][15]}

Tencent describes the engine as a "production-grade dual-system cognitive memory engine" for "self-evolving LLM agents".^[1] It keeps separate memory pipelines for three kinds of work: personal conversation (Chat), general productivity tasks (Working) and software engineering (Coding).^[1] The first version, Hy Memory 1.0, was announced on May 28, 2026 as a memory plugin for OpenClaw-style agents.^[8][9]

Overview

Tencent's launch page says the three layers can be used on their own or combined into a single agent memory stack.^[6]

History

Hy Memory 1.0

Tencent Hunyuan announced Hy-Memory on its WeChat account on May 28, 2026, describing it as a memory plugin built for long-running collaborative agents such as OpenClaw.^[8] The post framed the problem through what it called a "three-week trajectory", in which users stop trusting an agent for deeper work once it starts forgetting decisions made in earlier sessions.^[8] The 1.0 design rested on three ideas: a layered memory structure (six layers in the 1.0 description, from raw traces and atomic facts up to mental models and prospective intentions), a split between a fast "System 1" writer and a slower background "System 2" process, and "evolution chains" that link a revised memory to the memory it replaced instead of overwriting it.^[8] Installation went through OpenClaw's plugin system.^[8]

Tencent claimed in the announcement that Hy-Memory scored 85.2 on LongMemEval and beat other memory frameworks on LongMemEval and PersonaMem, cut the number of stored memories by more than 70%, raised information density per memory by more than 45%, lowered token use on very long contexts by 35% and sped up memory updates by 20%.^[8] These were the company's own figures. The tech news site AIBase reported the launch the following day, repeating the same claims.^[9]

The hy-memory package history on PyPI starts with version 1.0.0 on May 21, 2026, and the 1.x line ran to version 1.2.21 in July 2026.^[10] Tencent's 2.0 launch page nevertheless lists 1.0 as not open-sourced and released as "paper and product", and says the 1.0 design is described in arXiv paper 2606.09483, where it is called DCPM.^[6]

Hy Memory 2.0

Tencent posted the 2.0 launch on X on October 16, 2026, describing it as an "open-source model, framework, and MemHub" and "your private memory steward".^[7] The GitHub repository was created on October 15, 2026.^[1] Besides the May 2026 release of Hy Memory 1.0, its news section lists four newer items: the hy-memory==2.0.2 release and open-sourcing of the monorepo with MemHub; the Hy-Memory-30B-A3B weights and technical report; the CMB coding-memory benchmark paper in September 2026; and acceptance of the DCPM paper at EMNLP 2026, as stated by the repository, in August 2026.^[1] Version 2.0.1 reached PyPI on October 15, 2026 (UTC) and 2.0.2 on October 16, 2026 (UTC).^[10] According to the changelog, 2.0.2 carried SDK fixes and Apache-2.0 package metadata; the repository adopted the Apache-2.0 license text after 2.0.1 was published, and PyPI metadata for a published release cannot be changed.^[13]

Compared with 1.0, Tencent says 2.0 upgrades chat memory from scattered facts to event-centered timelines and adds the Working and Coding memory scenes. It calls Hy Memory "the industry's first" memory framework to propose working memory and code memory, a claim it makes for itself.^[6]

The Hy Memory engine

Scenes

The SDK routes every add() and search() call through a scene router before it reaches one of two pipelines.^[1]

When no scene is given, writes without tool calls go straight to chat with no LLM call, and writes containing tool calls are classified once into one of the three scenes. Searches in the default mode go to chat without an LLM call; an "agentic" search mode adds scene classification, query rewriting and time-range parsing.^[1]

Chat memory hierarchy

The Chat pipeline abstracts raw conversation upward through a numbered hierarchy.^[1]

Three operating modes trade cost against depth. lite writes and searches raw embeddings only, with no LLM calls, which amounts to plain vector retrieval-augmented generation. pro adds LLM extraction and reconciliation, with an optional graph store. ultra adds System 2 schema induction and a cross-domain "sweeper", triggered by the evolve() call, and always uses a graph database.^[1]

Working and Coding memory

For productivity work, each session that contains tool calls produces two linked layers at once: exactly one fact record describing what the session did (task, approach, result) and zero or more "experiences", reusable approaches that must pass a value threshold and carry fields such as the conditions under which they apply and a confidence score. Facts and experiences point at each other, and superseded facts stay traceable through a task-progress chain.^[1]

Coding uses the same two-layer pipeline with a software-engineering extraction spec that records exact coordinates (file, function, branch) rather than pasting code, and it stores data separately from Working memory.^[1] On top of this it keeps a version-evolution chain that stores the real diff of every revision of a block of code without any LLM calls. The README's reasoning is that fact extraction deliberately leaves code out, so a version that was later reverted would otherwise be lost, since it is usually no longer in the working tree either. The chain is attached under the facts that vector retrieval returns, rather than acting as a retriever itself. Its change sources are local Git diffs or explicit replace actions in the tool trace, and each chain thread covers one block of one function in one file on one branch.^[1]

Backends and API

The SDK needs only an LLM endpoint and an embedding endpoint, and any OpenAI-compatible service works.^[1] Supported storage is listed below.^[1]

The client exposes add, search, get, update, delete, list_memories and evolve. A source parameter (for example "openclaw", "cursor" or "codex") keeps memories from different calling applications apart, and searches can include or exclude sources.^[1] The package requires Python 3.9 or later.^[10] An OpenClaw plugin ships in the repository's plugins/ directory.^[1]

Research papers

DCPM (Chat memory)

The Chat design is described in "Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents" by Tianxiang Fei, Mingyang Song, Mao Zheng and Xiang Yu of Tencent's Large Language Model Department, posted to arXiv on June 8, 2026.^[4] The paper calls the system DCPM. A synchronous "daytime" writer (System 1) persists raw input, extracts facts and identity items, and reconciles each against nearby stored items with ADD, SUPERSEDE or UPDATE operations. On SUPERSEDE it keeps the old item and links old and new with pointers in both directions, forming a chain of belief revisions. An asynchronous "nighttime" engine (System 2) clusters new facts, uses an LLM agent to induce schemas and intentions in a graph, and runs a cross-domain sweeper that looks for schema pairs that are behaviorally similar but semantically distant and abstracts them into higher-level "core schemas". The read path uses no LLM calls.^[4] The paper borrows the System 1 and System 2 split from dual-process accounts of human cognition.^[4]

The authors evaluated DCPM on LongMemEval, PersonaMem and PersonaMem-v2 with DeepSeek V4-Flash and Kimi 2.5 as the extraction and inference models, against a long-context baseline, Mem0 and Zep/Graphiti.^[4]

The paper's argument is that System 2 should matter most where answers depend on implicit inference across sessions; enabling it added the most on PersonaMem-v2 (up to 5.20 points) and the least on recall-heavy LongMemEval.^[4] The full-context baseline still beat DCPM on several columns, including LongMemEval with DeepSeek.^[4]

CMB (Coding memory)

The Coding scene is tied to "CMB: A Coding Memory Benchmark for Long-Horizon and Interleaved Coding Agents" by Jiazheng Kang, Tianxiang Fei and Mao Zheng of Tencent Hunyuan, which the repository says was submitted to ICLR 2027 and hosts as a PDF.^[1][5] CodeMemoryBench (CMB) replays frozen, interleaved coding histories in real repositories to every memory method, then asks diagnostic questions and has the agent make graded code edits. It contains 216 code-editing tasks and 750 diagnostic questions over 5,595 turns from nine repositories in four languages.^[5]

The paper proposes a three-layer coding memory called TriLink (fact, experience and diff memory). Using Claude Sonnet 5 and DeepSeek V4-Flash as agent backbones, TriLink resolved 43.1% and 35.6% of CMB tasks, against 36.6% and 28.2% for the strongest existing memory baselines (Mem0, MemOS, RAG, OpenViking and RepoMemory were tested). On the external SWE-ContextBench suite it reached 47.5% and 28.3%, against 45.5% and 25.3%.^[5] A full-context reference with Claude Sonnet 5 resolved 58.8% of CMB tasks, above every memory method.^[5]

Hy-Memory-30B-A3B

Purpose and architecture

Hy-Memory-30B-A3B is a "compact long-term-memory writer model" that replaces the general-purpose LLM inside a memory framework's write path. It decides what to keep, how precisely to record it and how new evidence should update the stored bank.^[2] Despite the shared "30B-A3B" naming with Qwen3-30B-A3B (which appears only as a comparison baseline), the technical report says the model was trained from Tencent's own Hy3-30B-A3B instruction-tuned checkpoint, with the flagship Hy3 model as distillation teacher, and the published configuration uses Tencent's hy_v3 architecture.^[3][14]

The model card's quickstart serves it with vLLM (tested with v0.31.0) behind an OpenAI-compatible API, with a "no_think" reasoning setting for extraction.^[2]

Training

The technical report, "Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing", was posted to arXiv on October 15, 2026 by Tianxiang Fei, Jiazheng Kang, Mao Zheng, Xingyan He and Xiang Yu of Tencent Hunyuan.^[3] Training ran inside the real, unmodified write paths of three memory frameworks: Mem0, MemOS and Hy Memory (in its Chat, Working and Coding configurations).^[3]

The first stage is a form of knowledge distillation aimed at teaching each framework's output protocol, so that RL does not waste samples on malformed JSON or invalid operations.^[3] The report's main contribution is the reward design for the second stage, which tries to solve the credit-assignment problem in memory writing: a write made early in a conversation may only matter when a much later question asks about it.^[3] The reward combines three parts:

• Per-chunk recall reward. Probe questions test whether facts introduced in a chunk remain answerable after retrieval, and "trap" probes check that content which should not be stored (greetings, generic assistant encouragement) stayed out of memory.^[3]
• Attributed outcome reward. Cross-chunk questions come with gold facts that map recovered evidence back to the chunks that introduced it, so credit is not spread across every write in a sequence.^[3]
• Bank-growth penalty. Growth in stored tokens relative to input tokens is penalized above a threshold, to stop the policy from storing everything.^[3]

Advantages are normalized among rollouts at the same chunk position, so one rollout can be rewarded at one chunk and penalized at another. At the end of each window, the best-scoring bank is carried forward as the starting state of the next window, so training sees banks the policy itself produced.^[3] The final reward dataset holds 164,576 positive probes, 8,599 traps, about 60,800 outcome questions and 222,395 chunk-level gold-fact references, built over 12,671 windows and 57,616 chunks.^[3] Answer and judge models are held frozen so that only the memory bank changes.^[3]

Evaluation

Tencent's evaluation swaps only the memory writer and holds the framework version, prompts, retrieval settings, answer model and judge fixed. Each framework score is the equal-weight average of six metrics (LoCoMo judge score, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and CMB), all run on a fixed 20% subsample of each benchmark. EverMemOS was held out from training to test transfer to an unseen framework.^[2][3]

Source: Tencent's technical report and model card; self-reported results on 20% subsamples.^[2][3]

Tencent's launch post summarized this as "Hy-Memory-30B-A3B beats Opus 5 across major frameworks".^[7] The report itself is more careful. It says the 0.01-point and 0.02-point margins over Claude Opus 5 on MemOS and Hy Memory "indicate parity with Opus 5 under this deterministic sampled evaluation, not a statistically resolved difference", and that the stronger evidence is the consistent ranking and the 2.86-point lead on the held-out EverMemOS framework.^[3] The overall lead over Opus 5 is 0.87 points, and the gain over the Hy3-A3B base checkpoint is 7.84 points.^[3] The per-benchmark tables show the model does not win every metric: Claude Opus 5 scored higher on LoCoMo in three of the four frameworks (tying in the fourth) and on LongMemEval in three of the four, while Hy-Memory-30B-A3B's largest gains came from Travel SPS and CMB.^[2][3]

The report lists its own limitations: no controlled ablations of the reward components, no repeated training seeds, a single held-out framework, and attribution that is "evidence-based rather than counterfactual".^[3]

MemHub

MemHub is described as a "local-first" shared-memory product: it gathers history and new conversations from AI applications the user has authorized into one memory store, which those applications can then query after being separately authorized.^[11] It is built on the Hy Memory SDK, and the repository describes it as unifying conversation history and coding traces from tools such as Claude Code, Cursor, Codex, OpenCode and Qwen Code.^[1] MemHub's internal module, data directory and legacy command name are "One Memory" (one-memory).^[11]

Integration has two independent switches. One installs recall into a tool (typically via the Model Context Protocol plus a skill), and the other controls whether MemHub reads that tool's sessions. Turning sync off does not uninstall the recall integration or delete memories already created.^[12] The support matrix in the repository, dated October 14, 2026, lists ten tools.^[12]

Source: MemHub harness support index.^[12]

The launch tweet said MemHub "syncs across 10+ AI tools" and comes as a CLI and Mac app.^[7] Tencent's website lists the same ten tools, while the repository's own matrix marks eight of them as open.^[6][12]

Framework benchmark claims

Tencent's 2.0 site claims that the Hy Memory framework placed first among open-source memory frameworks on four benchmarks, one or two per scene, compared against Mem0, MemOS, EverOS, OpenViking, RepoMemory and a simple RAG baseline. The tweet's "#1 open-source on 4 benchmarks" refers to these results.^[6][7] Tencent states that these runs used DeepSeek V4.1 Flash as the memory LLM and GPT-5.6-sol as the answer and judge model.^[6]

Source: Tencent Hy Memory 2.0 website.^[6]

On MemoryArena the site says Hy Memory led the next-best framework, EverOS, by 10.2 points, while its stored memory amounted to 6.06% of input tokens.^[6] The CMB and SWE-ContextBench figures match the DeepSeek V4-Flash results for TriLink in the CMB paper.^[5][6]

Use inside Tencent

The launch site says Hy Memory is deployed in several Tencent products, and reports internal evaluations that have not been independently checked.^[6]

Source: Tencent Hy Memory 2.0 website.^[6]

Context

Hy Memory belongs to a group of agent memory systems that store information outside a model's context window and retrieve it later, alongside projects such as Mem0, MemOS and MemGPT.^[3][5] Most of these frameworks put a general-purpose large language model behind framework-specific prompts to do the writing.^[3] Hy Memory's two distinctive bets are visible in its papers: keeping superseded memories linked rather than overwritten, and training a dedicated small model for the write step that can be dropped into other frameworks.^[3][4]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22Tencent Hunyuan, "Hy-Memory" GitHub repository README. github.com/...Hy-Memory
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Tencent, "Hy-Memory-30B-A3B" model card, Hugging Face. huggingface.co/...Hy-Memory-30B-A3B
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23Fei, T., Kang, J., Zheng, M., He, X., Yu, X. "Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing." arXiv:2610.22196, October 15, 2026. arxiv.org/...2610.22196
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Fei, T., Song, M., Zheng, M., Yu, X. "Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents." arXiv:2606.09483, June 8, 2026. arxiv.org/...2606.09483
5. ^1 ^2 ^3 ^4 ^5 ^6Kang, J., Fei, T., Zheng, M. "CMB: A Coding Memory Benchmark for Long-Horizon and Interleaved Coding Agents." PDF in the Hy-Memory repository. github.com/...on_and_Interleaved_Coding_Agents.pdf
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Tencent Hunyuan, "Hy-Memory 2.0" official website, October 16, 2026. memory.hunyuan.tencent.com
7. ^1 ^2 ^3 ^4 ^5Tencent Hy (@TencentHunyuan), "Hy Memory 2.0 is out" post on X, October 16, 2026. x.com/...2111027329816465435
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7Tencent Hunyuan, "Hy-Memory 发布：打造记忆力超强的 Agent 第二大脑" (Hy-Memory release), WeChat, May 28, 2026. mp.weixin.qq.com/...dH5PGpxqArFZkSq3-QxWYA
9. ^1 ^2AIBase, "Memory Capacity Increased by 45%, Token Cost Reduced by 35!! Tencent Hunyuan Launches the Strongest Memory Plugin Hy-Memory," May 29, 2026. news.aibase.com/...28443
10. ^1 ^2 ^3 ^4"hy-memory" project page and release history, PyPI. pypi.org/...hy-memory
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Tencent Hunyuan, "MemHub / One Memory" README. github.com/...README.md
12. ^1 ^2 ^3 ^4Tencent Hunyuan, MemHub harness support index (updated October 14, 2026). github.com/...README.md
13. ^1 ^2Tencent Hunyuan, Hy-Memory CHANGELOG. github.com/...CHANGELOG.md
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Tencent, Hy-Memory-30B-A3B `config.json`, Hugging Face. huggingface.co/...config.json
15. ^1 ^2Tencent Hunyuan, Hy-Memory LICENSE file. github.com/...LICENSE