Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments
3.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.19128.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryPaper extends the SwiftSage dual-process language agent with two modular components: an Adaptive Memory Module (salience-gated episodic storage with trigger-driven retrieval) and a Self-Reflection Module (bounded execution-time validation). Controlled ablations on ScienceWorld across four configurations show the full system scores best; SRM is the strongest standalone contributor.
Why it mattersIdentifies execution-time control as the dominant bottleneck in interactive language agents, with episodic memory only paying off once the runtime loop is stabilized — a useful framing for agent design decisions.
Cited by
No citations on record.