Liang Wenfeng's leaked four-hour memo, read through our 400-turn memory test

The provenance first. These are leaked notes from a four-hour closed-door investor meeting with DeepSeek founder Liang Wenfeng. Fragments circulated in Chinese AI and investor circles for about a month before Sina Finance compiled them on July 23, 2026, crediting an institutional research note. DeepSeek has confirmed none of it. Every quote below is translated from that compilation, and every judgment leans only on numbers we ran ourselves.

The headline items have already been repeated everywhere: the five things DeepSeek won't do (no “genius” mythology, no profit maximization, no closed source, no blind fight for users, no video generation, 3D, or world models, and no attempt at the next super-app), plus the line “restraint is a strategy: giving things up raises the probability of reaching AGI” (translated). You don't need us for a recap. What we can add is a translation layer: what this meeting means if you use AI to continue novels, run roleplay, or keep a companion. The reason to write it is simple. The heaviest sentence in the memo lands squarely on an exam we happened to grade two days before it surfaced.

The “continuous learning” he demands is the thing our test found missing

Per the notes, Liang lays the road to AGI out as a staircase: last year's step was chain-of-thought, this year's is agents, the next is continuous learning, and beyond that AI iterating on itself. He calls the whole climb a “gradual singularity” (translated). On the continuous-learning step his wording is the strongest in the room: humans learn on the job, but an AI has to be handed the full context all over again for every new task, which he calls “almost impossible” (translated). Then the sentence now being quoted everywhere: “the next generation of models must have continuous-learning capability; otherwise it cannot be called a next-generation model” (translated).

We had just measured the defect he is describing. On July 21 we ran a 400-turn companion-chat memory exam: 50 memory probes planted in everyday small talk, 27 left after voiding any probe whose answer leaked into nearby turns, three models examined the same day under one protocol: deepseek-v4-pro, qwen3.8-max-preview, qwen3.7-max. First pass, bare context: the model sees only the last 16 turns, a stand-in for a chat app with no memory system. All three flunked together: 19%, 22%, 19%.

The whole spread is one question. A newer model didn't help. A different vendor didn't help.

Second pass, we injected 8 retrieved memory entries into the prompt. Same three models: 89%, 74%, 59%. The models didn't change; what changed is whether someone handed them the facts. Liang is describing task-level amnesia at work; we measured turn-level amnesia in chat. Same disease. Weights don't update with use, so whatever isn't re-sent into the context may as well never have happened. His bar and our table fit together neatly: the founder defines continuous learning as the entry ticket to the next generation, and the current generation starts at 19% without app-layer help.

He isn't just talking, either. The Engram paper from January 2026, with Liang on the author list, already shows one path: a conditional-memory module that moves static knowledge out of neural compute into O(1) table lookups, lifting long-context retrieval accuracy from 84.2 to 97.0 in the paper's benchmark. That is memory at the architecture layer. Retrieval injection is memory at the app layer. Different floors of the same building, same direction: remembering should stop requiring recomputation.

Until then, memory engineering is not optional

No timeline appears in the notes. Before models learn to remember on their own, “the AI remembers you” has exactly two implementations. Resend hundreds of turns of history every turn and watch the bill amnesia arrive before the model's does; the arithmetic is in our context-window explainer. Or have the app retrieve the relevant facts and inject them, which is the entire meaning of that 19-to-89 jump. Foreverse bets on the second: tavern chats run a memory plugin with a fact store plus retrieval, and companions keep a memory profile you can read, edit, and delete entry by entry. For users, the founder's bar translates into one shopping rule: until the next generation actually ships, judge an app by its memory system before you judge it by its model.

“We only make a reasonable profit” — so will DeepSeek raise prices?

The pricing passages are worth reading slowly. “We only make a reasonable profit” (translated). When V4 launched, the team worried demand would outrun supply and priced high; after cutting to one quarter, “many people in the company group chat cheered” (translated). “Low cost is a result of the architecture” (translated). Ranking competitive factors, he puts cost above time and user experience: “cost ranks first” (translated). He even sounds lukewarm about the business itself: “I don't find selling APIs all that attractive” (translated), and “we have always been commercializing, we just don't treat commercialization as the goal” (translated).

Read that alongside what the company did on June 29: peak-hour pricing, announced by email to all API developers. All billing items double during 9:00–12:00 and 14:00–18:00 Beijing time; the other 17 hours keep the post-cut rate. So “reasonable profit” evidently includes using price as a load-balancing lever. The philosophy constrains the direction, not the number. Prices moved three times in 2026, in both directions. If you write or read with DeepSeek in the evening (off-peak in Beijing covers most Western daytime hours), the doubling barely touches you; batch jobs are worth scheduling. For the broader picture of where token prices are heading, we keep a running price-trend log. Budget from the official pricing page, not from anyone's retelling, including ours.

Still waiting for a DeepSeek image or video model? The memo says stop

The most practical item in the five-no list is the least discussed. The translated line leaves little room: “many things are not on our main line, including 3D and video generation.” World models, per the notes, don't have much to do with the ceiling of intelligence. Multimodality gets a nod as important for products and consumers, but “not the main line, nor intelligence itself” (translated).

The operational reading takes one sentence: if your plan was to wait for DeepSeek to ship image or video generation, retire the plan. Illustrating a chapter, giving a character a voice, cutting a scene to video — the long-term shape is one model per job. Prose goes to a text model chosen on continuation test results, images to an image model, speech to TTS. Foreverse configures text, image, speech, and video models separately for exactly this reason.

What we won't vouch for

Once more for the record: every quote here comes from an unconfirmed leak, and an official correction could land tomorrow. We used it because the checkable parts check out — the quarter-price cut matches the mid-year announcement, peak pricing matches the June 29 email, continuous learning matches the January Engram paper. The passages about team and fundraising we could not verify, so they don't appear in this post. As for when continuous learning actually arrives, we'll do what we did in the exam post: leave a dated stake in the ground. July 2026, bare context 19%, injected 89%. The day a model ships wearing the “continuous learning” label, the same 400-turn exam paper is ready for it.