AI TrackerFrontier desk · AEST
Back to news
normalaitech

Compaction, decay and reinforcement in a two-tier conversational memory: a paired evaluation against retrieval-augmented generation

arXiv cs.CL · 7 Nov 2026, 03:56 AEST

A study compares two methods for conversational agents: retrieval-augmented generation and compaction of dialogue into atomic facts. The results show that compaction outperforms retrieval-augmented generation in certain scenarios, particularly when answers are factual rather than phrased.

Key facts

Read aloud

A study compares two methods for conversational agents: retrieval-augmented generation and compaction of dialogue into atomic facts. The results show that compaction outperforms retrieval-augmented generation in certain scenarios, particularly when answers are factual rather than phrased. Retrieval-augmented generation keeps the transcript and searches it; the alternative is to compact each turn into atomic facts and search those. On the four categories comparable work reports, compaction scores seventy one point four per cent against sixty four point six, a lead of six point eight points p = five.5e-eight at thirty four per cent fewer prompt tokens. Our lead falls as the reference answer demands more source wording, nine point six points ahead at one or two reference words and three point one behind past eight. Compaction keeps what a fact asserts and discards how it was said, so it wins where the answer is a value and loses where it is a phrasing.

Sources in this cluster