UniMem: complementary routing from episodic to parametric memory¶
Fidelity: Runs the paper's isolatable state and update mechanisms on common public data or a deterministic mini-suite. This does not reproduce the original large model.
Paper information¶
| Field | Details |
|---|---|
| Paper | UniMem: complementary routing from episodic to parametric memory (arXiv 2607.26017) |
| Organization | CASIA / UCAS / Peking University / University College London |
| First public release | 2026-07-28 |
| Authors' code | No official public repository was found as of 2026-07-29. |
| Adapter / method | unimem |
| Local reproduction code | src/auto_research/agent_research/ |
Original paper¶
Motivation and method¶
Novel tasks first enter an episodic buffer. A self-routing controller consolidates recurring and reliable execution patterns into scalable parametric memory.
flowchart LR
X["Unbounded task stream"] --> R["Routing token"]
R --> E["Episodic buffer"]
R --> P["Parametric memory"]
E --> C["Consolidate reliable patterns"]
C --> P
Key figure from the paper¶
Figure 1: Overall workflow, stages, and data flow. From the original paper; copyright remains with the authors.
Core equations¶
Paper results¶
On long-horizon streaming tasks, execution fidelity is maintained and exact match improves by 4.0 points on average across three backbones. No production online A/B is reported; only offline or environment outcomes are included.
Local reproduction¶
The EvoMem mini-suite runs episodic routing, the recurrence threshold, parametric routing, and consolidation statistics.
| Metric | UniMem |
|---|---|
| Joint success | 1.0000 |
| Average cost | 0.5200 |
| Episodic / parametric / consolidation | 24 / 96 / 12 |
See the fixed-seed metrics.
Reproduction boundaries¶
Parametric memory is a structured pattern table, not expanded LLM weight blocks. The run checks the stability/plasticity routing logic.
