AI can’t remember you? Volcano engine open source OpenViking context database actual test｜Web version requires no installation

Original author：David Yin 不确定体验局

Original source：https://www.youtube.com/embed/iluN3447mPE

AI can't remember what you said - Volcano Engine has open sourced a set of tools for "installing memory for AI" called OpenViking. You can try it by opening the browser without installing or logging in. I clicked all the way to the memory that said "Keep your answer short" in the official Studio, and there was only one sentence in English when I clicked on it; I also truthfully reported that it was not easy to use today. This issue explains clearly: why AI forgets you as soon as the dialog box is changed; after OpenViking arranges the memory into a `viking://` virtual file system, the memory becomes translatable, readable, and changeable; how the three-layer on-demand loading (one-sentence summary/one-page overview/original text) eliminates useless context; how the session is extracted into memory (new/merge/ Skip); and two pitfalls - the main project is AGPL-3.0 (used for external network services, your changes must also be open source), self-deployment requires Python 3.10+ and you must prepare your own Embedding model and VLM. Official self-test data (**Official self-report, not third-party retest, tested on OpenViking 0.3.22, Embedding and VLM both use Beanbao**): Long conversation memory accuracy OpenClaw 24.20% → 82.08%, Hermes 33.38% → 82.86%, Claude Code 57.21% → 80.32%; enter token It is reduced by 34.3%–91.0%, and the query delay is reduced by 58.45%–66.10%. My own actual measurement boundary: I searched three times in **Public Demonstration without Login**, and found 0 / 0 / 1 hits in the memory. The only hit was that the abstract was not the original text - there was only one sample, no login, and the index status was unknown, so this is only an actual measurement record, not a conclusion about the product. Chapter: 0:00 Why does AI forget you as soon as the dialog box is changed? 0:46 The official one-sentence definition for OpenViking 2:09 Is it free/do I need to install it/do I need a graphics card? 2:33 What do I see when I open the webpage? 3:14 How is the memory saved: viking:// virtual file system 3:54 Click on the file, and there is only one sentence in it 4:26 How to find too much content: three layers of on-demand loading 5:08 I searched three times, but not a single memory was found twice. 5:36 AGPL-3.0 This pitfall 6:25 One thing you can do today. First-hand source: OpenViking warehouse https://github.com/volcengine/OpenViking Official Studio (no installation required) https://openviking.ai/studio Official Chinese documentation https://docs.openviking.ai/ Context layering (L0/L1/L2) https://docs.openviking.ai/zh/concepts/03-context-layers My own website (past AI vernacular videos and long articles are archived here according to time, and they can be found faster than browsing on the platform): https://davidyin.dev https://davidyin.dev/writing/ https://davidyin.dev/video/ Subscribe to the channel 🔔 Don’t miss the next issue of #AIMemory #OpenViking #OpenSourceAI