What this is
ByteDance's open-source framework DeerFlow breaks open how Agent long-term memory is implemented, and the conclusion runs counter to intuition: by default, no vector database (a database that retrieves by semantic similarity) is used—the LLM proposes, the code decides, and the simple approach is actually more reliable.
Its Memory design runs in three steps:
- Filter: Strip out tool calls, acknowledgments ("OK," "thanks"), temporary files, and other clearly valueless messages;
- Propose: A dedicated Memory LLM evaluates "is this worth remembering," outputting structured facts;
- Adjudicate: A code-layer check—must be user-level, long-term valid, with confidence ≥ 0.7, before anything is actually written to disk.
The design also introduces debounce (triggering only after the user pauses input) and watermark (a marker tracking processed position) to control duplicate calls.
Industry view
Supporters say this is the real watershed for Agent productization: whether users feel the AI "gets me" hinges on memory quality, not model size. Fine-grained filtering and adjudication turn a general-purpose model into a personal assistant.
Skepticism exists too. JSON file storage means losing everything when switching devices—cross-device sync is a major hole. The extra Memory LLM call per round adds latency and cost that can't be ignored. More importantly, while this works in single-user scenarios, when 1,000 employees run concurrently, no one has answered how to arbitrate conflicting facts or draw privacy boundaries. Some voices also point out that DeerFlow demonstrates a direction, not an end state—when Agents enter enterprise core workflows, memory system complexity will climb another order of magnitude.
Impact on regular people
For enterprise IT: When evaluating AI assistants, "does it remember me" will matter as much as "is it smart." The selection checklist needs a new line item: long-term memory mechanism.
For working professionals: The AI tools you use daily will increasingly split into two camps—"ones that get you" and "ones that don't." The difference often isn't the model, but the memory design behind it.
For the consumer market: AI companion and assistant products will further segment. Products that remember have higher stickiness, but user demands for data ownership and the "right to be forgotten" will intensify in parallel.