AuthorsJingbo Ji, Lingyi Li, Xilong Cheng et al.
TL;DR
RippleMem uses cue-rich episodic memory units plus anchor-local associative recollection to reach 87.14% judge accuracy on LoCoMo, +3.31 points over RF-Mem.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
RippleMem highlights that answer-critical evidence can be stored yet missed, as flat retrieval returns isolated records and undirected association surfaces non-supporting memories.
In long-horizon conversational agents, this causes queries that depend on distributed events, preferences, or constraints to be answered incorrectly despite the relevant history being present.
HOW IT WORKS
RippleMem’s core mechanism combines Cue-Rich Episodic Memory Construction, an Event-Centric Memory Graph, Adaptive Associative Recollection, and Evidence Assembly to turn dialogue into cue-rich event memories.
You can think of RippleMem like a hippocampus-backed card catalog: events are indexed by semantic vectors and episodic cues, then recalled and expanded from anchors when partial evidence is found.
This associative recollection lets RippleMem recover missing support across sessions and time, something a plain context window or one-shot RAG cannot reliably achieve.
DIAGRAM
This diagram shows how RippleMem processes a query through hybrid initial recall, anchor planning, graph expansion, and evidence assembly during inference.
DIAGRAM
This diagram shows how RippleMem is evaluated on LoCoMo and LongMemEval-S, including baselines and ablation variants.
PROCESS
Cue-Rich Episodic Memory Construction
RippleMem processes dialogue windows into schema-guided memory units with restatements, dense vectors, and episodic cues Pi, Li, Ti, ensuring each event is independently retrievable.
Event-Centric Memory Graph
RippleMem links memory units into a sparse weighted event graph using semantic association ssem and structural association sstr over participants, locations, and temporal intervals.
Adaptive Associative Recollection
Given a query, RippleMem performs hybrid initial recall, then uses a recollection controller to select anchors Ar and missing-support target gr, expanding locally in the graph to recover support Sr.
Evidence Assembly
RippleMem consolidates provenance, scores memories with ρ(m) combining alignment and anchor status, selects top K units as EK, and passes them to the answer generator for grounded responses.
KEY CONTRIBUTIONS
Evidence Recovery Formulation for Long-Term Memory
RippleMem frames long-term memory as recovering an answerable evidence set from distributed traces, highlighting that one-shot query matching can miss stored support across interactions.
RippleMem Event-Centric Associative Recollection
RippleMem combines Cue-Rich Episodic Memory Construction, an Event-Centric Memory Graph, and Adaptive Associative Recollection so recalled memories act as both context and cues for further support.
Benchmark Gains with 30× Cheaper Graph Construction
RippleMem achieves 87.14% judge accuracy on LoCoMo and up to 86.60% on LongMemEval-S, while reducing graph construction time by about 30× compared with Zep on LoCoMo.
RESULTS
LLM-as-a-Judge Accuracy
87.14%
+3.31 over RF-Mem on LoCoMo
F1
52.49
+1.01 over SimpleMem on LoCoMo
BLEU-1
44.05
+1.54 over SimpleMem on LoCoMo
LongMemEval-S Overall Accuracy
86.60%
+3.60 over EverMemOS in its evaluation setting
On LoCoMo, a long-term conversational memory benchmark, RippleMem’s 87.14% judge accuracy and 52.49 F1 show more complete evidence recovery than RF-Mem and SimpleMem. On LongMemEval-S, RippleMem’s 86.60% overall accuracy demonstrates strong multi-session reasoning and temporal tracking in full-history user agent interactions.
BENCHMARK
On LoCoMo, a long-term conversational memory benchmark, RippleMem’s 87.14% judge accuracy and 52.49 F1 show more complete evidence recovery than RF-Mem and SimpleMem. On LongMemEval-S, RippleMem’s 86.60% overall accuracy demonstrates strong multi-session reasoning and temporal tracking in full-history user agent interactions.
BENCHMARK
LLM-as-a-Judge accuracy on LoCoMo across RippleMem and key baselines.
BENCHMARK
Overall LLM-as-a-Judge accuracy on LongMemEval-S under the EverMemOS evaluation setting.
KEY INSIGHT
RippleMem cuts graph construction time on LoCoMo from 3532.03 seconds for Zep to 117.51 seconds, roughly a 30× reduction while increasing judge accuracy from 76.69 to 87.14.
This is counterintuitive because richer associative recollection and event graphs usually seem more expensive, yet RippleMem’s sparse, incremental design yields both higher accuracy and dramatically lower construction cost.
WHY IT MATTERS
RippleMem unlocks long-term agents that can treat initial recall as a starting point, using anchors and episodic cues to complete missing evidence across sessions and timelines.
Builders can now design memory systems where stored experience is organized into cue-rich events and recollection is guided by evidence needs, making reliable long-horizon reasoning practical without massive context windows.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…