UniMem：LLM Agent 终于不用在「记太多」和「学不会」之间二选一了

When an LLM agent runs in production long enough, a pattern emerges. External retrieval-based memory absorbs new information quickly but never fully internalizes recurring execution patterns, and every inference carries retrieval overhead. Parametric memory — the weights inside the model — is fast once something is learned, but depends on explicit task boundaries, a fixed parameter budget, and retraining processes that don't suit continuous deployment. This is the stability-plasticity dilemma.

A new paper by Siyu Xia and ten co-authors proposes UniMem, a self-routing framework that manages memory autonomously. Novel or sparse experiences stay in episodic memory — detailed, recallable, retrieval-augmented. Recurring patterns, once they prove reliable, get consolidated into parametric memory — automatic, zero retrieval cost. UniMem uses learnable routing tokens as memory controllers, decoupling task identification from execution and deciding on the fly which pathway a task belongs to.

During deployment, no task labels are needed. Parameter bloat stays controlled, and memory expands on demand. On long-horizon streaming task sequences, UniMem outperformed baselines while maintaining execution fidelity, averaging a 4.0 Exact Match points gain across three backbone models.

Open questions remain. How far can the expandable parametric memory scale before hitting diminishing returns? Do the routing tokens generalize across very different task domains? And 4.0 EM, while consistent, leaves room to grow. Still, for teams building agents that need to run continuously, handle evolving task streams, and improve over time without manual babysitting, this architecture is worth examining.

相关链接：
- 论文页面：https://arxiv.org/abs/2607.26017
- PDF 全文：https://arxiv.org/pdf/2607.26017