Tencent releases open-source Hy Memory 2.0 with 30B model; test finds quality hinges on the writer LLM

Tencent has released Hy Memory 2.0, an open-source effort to give AI assistants memory that persists across sessions and tools. The release bundles three parts: a framework under the Apache-2.0 licence, a 30-billion-parameter memory model, and MemHub, an application that lets multiple AI tools read from and write to a single shared memory.

How Hy Memory stores what it learns

According to a dev.to analysis published this week, Hy Memory uses a language model to compress each conversation into discrete facts and reconcile them with what was stored before; when a value changes, the old figure is kept as history next to the current one. The framework works with any OpenAI-compatible model, and its README points to DeepSeek. Tencent's own 30B model is not exposed through an API the author could access, but Tencent's model card reports an average of 55.5 for it against 54.6 for Claude Opus 5, which made Opus a defensible stand-in for testing.

The benchmark

The evaluation carries a caveat its author states plainly: they build Aura, a competing open-source memory library, though they also report the results that favoured their rival. The test used LongMemEval, a public benchmark in which each question must be answered from a user's history of roughly 50 conversations. Three of its six question types were used, 20 questions each: one probes whether memory returns the newer version of a fact that has changed, one where the answer exists only in the assistant's own earlier replies, and a control drawn from user statements. Gemini 3.8 Flash both read the retrieved memory and produced answers for both systems, and LongMemEval's official judge prompts graded the output.

Aura 1.62.0 takes the opposite approach to Hy: it stores every turn verbatim with its speaker and date, retrieves the ten most relevant, and calls no model at write time.

The writer model decides the score

With Gemini Flash writing Hy's memories, Hy answered 43% of the 60 questions to Aura's 95%. The assistant-statement category was the widest gap: 0% against 100%. Updated facts came in at 55% versus 90%, and the user-statement control at 75% versus 95%. Thirty-two of Hy's 34 misses were the answering model saying it simply did not have the information.

Retrieving more did not help. Re-searching Hy's store with a larger limit matched Aura's context length and still scored 45%; 40 memories, roughly 7,900 characters, also scored 45%. The author concludes the loss happened while conversations were being condensed into facts, not during search.

A second run used Claude Opus 5.5 as the writer. At roughly $5 per user history, only ten questions per type were affordable. There Hy matched Aura on updated facts, nine out of ten each, a result the author flags as going against their own product. On assistant statements it reached four out of eight against Aura's eight out of eight, with every miss again an information-not-available answer.

What accuracy costs

Indexing one 50-conversation history took Hy 84 model calls and about 416,000 input tokens with Flash as the writer, roughly 15 minutes. With Opus it took 92 calls, about 695,000 input tokens plus 61,000 output, near $5 at list price, and 11 minutes. Aura processed the same history in 36 seconds with no model calls at all.

The mechanism behind the losses, as the author describes it: Hy reduces conversations to statements about the user, discarding whatever the writer does not treat as a user fact, such as the assistant's recommendations, details that seemed minor, or occasionally the newer value of a changed fact.

Limits of the test

The author lists the constraints: only three of six LongMemEval categories, leaving multi-session, temporal and preference questions untested; small samples where each question is worth 5 to 12 percentage points; Tencent's actual model untested; only the default pro mode rather than the graph-based ultra mode; a single reader and judge model; and 1.3% of conversations that Hy silently failed to store.

Why it matters

Shared memory across AI tools is quickly turning into infrastructure, and this data suggests the deciding factor is not the memory framework itself but the model that writes to it. A cheap writer dropped Hy to 43% overall and zero on assistant statements, while a frontier-class writer matched a no-LLM approach at around $5 per user history. Assistant-side utterances, the kind behind questions like what did you recommend last time, look like a structural blind spot for systems that extract user-centric facts. Anyone evaluating a memory layer should test it with the exact writer model they would deploy, price that per user, and verify the assistant's own words survive the compression.