AuthorsSahil Sen, Elias Lumer, Anmol Gulati, Vamse Kumar Subbiah
TL;DR
Chronos uses dual calendars plus dynamic prompting and structured temporal event retrieval to reach 95.60% accuracy on LongMemEvalS, +7.67% over EmergenceMem Internal.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing conversational memory systems either overbuild global knowledge graphs or rely on shallow turn retrieval, failing on time-sensitive multi-session queries.
On LongMemEvalS, EmergenceMem Internal only reaches 86.00% accuracy, showing that long-term assistants still mis-handle temporal reasoning and knowledge updates across months of interaction.
HOW IT WORKS
Chronos centers on Event Extraction, Dynamic Prompting, Initial Retrieval, and the Chronos Agent, backed by a structured event calendar and raw turn calendar.
You can think of Chronos like a calendar plus diary: the event calendar is a timestamped index, while the turn calendar is the full narrative notebook.
By explicitly structuring temporal events while keeping full dialogue, Chronos enables precise time filtering and multi-hop reasoning that a plain context window cannot support.
DIAGRAM
Sequence of interactions between the user, Chronos Agent, and the dual calendars during query-time tool-calling.
DIAGRAM
Flow of evaluating Chronos configurations and ablations on the LongMemEvalS benchmark.
PROCESS
Event Extraction
Chronos runs the Event Extraction pipeline over conversation turns, producing subject verb object tuples with ISO 8601 datetime ranges and lexical aliases into the event calendar.
Dynamic Prompting
For each new question, Chronos uses Dynamic Prompting to analyze the query and generate retrieval guidance bullets describing targets and temporal constraints.
Initial Retrieval
Chronos performs Initial Retrieval over the turn calendar using dense search, Cohere Rerank v3, and context expansion around the top 15 turns.
Chronos Agent
The Chronos Agent runs a ReAct loop, calling vector and grep tools over the event calendar and turn calendar until it can answer with temporally grounded reasoning.
KEY CONTRIBUTIONS
Chronos Architecture
Chronos introduces dual event calendar and turn calendar stores plus the Chronos Agent, achieving 92.60% and 95.60% accuracy on LongMemEvalS under Low and High configurations.
Dynamic Prompting for Memory
Chronos extends Dynamic Prompting to long-term memory, generating per-question retrieval guidance instead of static query rewriting across temporal, preference, and aggregation tasks.
Structured Event Retrieval Gains
Chronos ablations show the event calendar yields a 58.9% gain over the baseline, while other components each add between 15.5% and 22.3% accuracy improvements.
RESULTS
Overall
92.60%
+7.67 over EmergenceMem Internal
Knowledge Update
96.15%
+12.82 over EmergenceMem Internal
Multi Session
91.73%
+10.53 over EmergenceMem Internal
Temporal Reasoning
90.23%
+4.52 over EmergenceMem Internal
Chronos is evaluated on the LongMemEvalS benchmark with 500 questions across six categories, testing knowledge updates, aggregation, and temporal reasoning. The 92.60% and 95.60% overall scores show Chronos handles long-horizon, time-grounded memory substantially better than EmergenceMem Internal and Mastra.
BENCHMARK
Chronos is evaluated on the LongMemEvalS benchmark with 500 questions across six categories, testing knowledge updates, aggregation, and temporal reasoning. The 92.60% and 95.60% overall scores show Chronos handles long-horizon, time-grounded memory substantially better than EmergenceMem Internal and Mastra.
BENCHMARK
Overall accuracy (%) on LongMemEvalS across practical conversational memory systems.
BENCHMARK
Overall accuracy (%) for Chronos High and strong baselines under advanced LLM configurations.
KEY INSIGHT
Removing the event calendar almost halves Chronos Low’s accuracy, dropping from 93.1% to 58.6% on the 116-question ablation subset.
This is surprising because many systems assume dense turn retrieval is sufficient, yet Chronos shows structured temporal events dominate gains while using relatively simple SVO tuples.
WHY IT MATTERS
Chronos unlocks reliable, time-aware conversational memory, letting agents answer questions like “What did I do the week after my vacation?” months later.
Builders can now design assistants that track evolving preferences, knowledge updates, and cross-session event counts without heavyweight knowledge graphs or full-context replay.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…