We've noticed that the open-source memory framework Mem0 recently completed a directional adjustment—the v3 version pulled its LoCoMo test score from 71.4 to 92.5. The jump is significant, but what concerns us more is the mainstream practice it abandoned: switching from "write-and-modify in real time" to "append-only, judge at query time."
What this is
Mem0 is an open-source framework that gives AI Agents (i.e., AI programs capable of autonomously completing multi-step tasks) long-term memory. The traditional approach (v2) works like this: when a conversation ends, the system first extracts new facts, then checks existing memories, and lets the model decide whether to add, update, or delete. The store stays clean, but a single erroneous deletion is irreversible.
v3 flipped the write strategy: when a user "lives in Beijing first, then moves to Shanghai," both records are kept. At query time the system blends semantic similarity, keyword matching, entity recognition, and temporal ordering to surface the current fact.
Industry view
Supporters see this as a more stable engineering choice: append-only preserves history, the "judgment" is deferred until it's actually needed, and it saves some model calls during the write phase.
But the skepticism is equally clear: every query now has to fuse multiple signals, and the engineering complexity is not low; the memory store keeps expanding, and long-term governance costs remain unknown. Mem0 itself cautions that v3's retrieval strategy still needs validation against longer contexts.
On the other side, the team behind Memobase is walking the opposite route: instead of a general-purpose pipeline, they focus on a dual-track of "user profile + event memory"—stable attributes go into profile slots, things that happened go on a timeline, and model calls are amortized through batching.
The design spaces these two routes cover don't overlap: one leans toward "general-purpose memory store," the other toward "vertical user profile." There's also cognitive science lurking in the background: human memory splits into episodic (specific events), semantic (general knowledge), and procedural (behavioral flows), and AI memory design is currently finding its place within this taxonomy.
Impact on regular people
For enterprise IT: when choosing AI assistants going forward, memory architecture will become a real selection dimension—"clean but fragile" and "complete but complex" are two different promises.
For individual professionals: the quality of the "impression" you leave through long-term collaboration with AI depends on whether it makes judgments at write time or at query time; the former is cheap but error-prone, the latter is accurate but expensive.
For the consumer market: the ability to "truly remember you" will become a differentiating selling point for AI products, but whether users are willing to pay the price for "being remembered" will determine how far this path actually goes.