Hy-Memory: Attribution-based reinforcement learning for long-term memory writing¶

简体中文

Research reference only; reviewed and not reproduced because of cost. Published model and framework entry points do not mean the repository completed 30B memory training.

Paper information¶

Field	Details
Paper links	arXiv 2610.22196; pinned v1 full text
Company / institution	Tencent Hunyuan; first author Tianxiang Fei
First public release	2026-10-15 16:46:47 UTC (arXiv v1; announced October 16)
Author code	Hy-Memory framework, Apache-2.0; checkpoint entry. Full training materials have not been confirmed public; model terms require separate review.
Local adapter / stable key	`hy-memory`, reference only, no executable adapter
Local implementation	None. This page and catalog metadata only.
Review status	Full experiment not reproduced; review complete; no local GPU receipt

Paper summary¶

Background and main changes¶

First use OPD on memory states written by the student to learn write protocols for different frameworks. Then freeze the reader and judge and optimize only the writer. Chunk-level probes supplement sparse rewards; cross-chunk factual attribution locates contributions; a growth penalty discourages storing everything.

Architecture and data flow¶

flowchart TD
  B[Parent memory store] --> W[Writer rollout in isolated copy]
  W --> M[Candidate memory stores]
  M --> R[Frozen reader answers]
  R --> J[Frozen judge and fact attribution]
  J --> C[Per-chunk GRPO]
  J --> S[Choose final candidate memory store]
  S --> B

Original paper figure¶

Paper Figure 1: Overall method and main components. From the original paper; copyright remains with the authors. Click the image to open its source.

Core equation¶

Section 6 uses a current-chunk probe, attributed outcome, and growth penalty:

\[ s_i[t]=\mathrm{prm}_i[t]+2\mathrm{orm}_i[t]-0.3\mathrm{pen}_i[t],\qquad A_{i,t}=\frac{s_i[t]-\mu_t}{\sigma_t+\epsilon}. \]

Each rollout contributes once per chunk, not once per model call. Gold belongs only to the judge and is not passed to writer, reader, or retrieval queries.

Offline and online results reported in the paper¶

The paper compares memory frameworks, including out-of-training EverMemOS, on a fixed 20% benchmark subset, not the full benchmarks. Differences after replacing writer, reader, and judge together cannot be attributed to writer training alone.

No online production A/B is reported. Releasing a model and framework does not establish user-traffic gains.

Local reproduction¶

Reference only. Section 7 uses 30B-A3B, 128 H20s, and batches of 64 windows × eight rollouts. No model download, multi-framework RL, or service evaluation sweep. Reviewed and not reproduced because of cost.

Reproduction limits¶

No executable adapter, training score, or Evolve promotion. Other memory operators are not a reproduction of this paper. See the batch acceptance contract.