💬 NLP

Chronos: Temporal-Aware Conversational Agents with Structured Event Retrieval for Long-Term Memory

The paper introduces Chronos, a novel temporal-aware memory framework that structures dialogue into event tuples and utilizes dynamic prompting with iterative tool-calling to achieve state-of-the-art performance on long-term conversational memory tasks.

Original authors: Sahil Sen, Elias Lumer, Anmol Gulati, Vamse Kumar Subbiah

Published 2026-03-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sahil Sen, Elias Lumer, Anmol Gulati, Vamse Kumar Subbiah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are talking to a very smart friend who has been chatting with you for months. You ask, "What was the name of that restaurant we went to right after I got my new dog last summer?"

A standard AI assistant might get confused. It has read all your chats, but it sees them as a giant, messy pile of text. It doesn't know which part of the pile happened "last summer," or that "getting the dog" is the key event that links the two. It might guess, or worse, make up a fake restaurant.

Chronos is a new system designed to fix this. Think of it as giving your AI assistant a super-organized digital planner and a smart librarian working together.

Here is how Chronos works, broken down into simple concepts:

1. The Two-Book Library (Dual Indexing)

Most memory systems try to read every single word of every conversation and store it all in one big book. This is slow and messy.

Chronos splits the memory into two distinct books:

The Storybook (Turn Calendar): This is the raw conversation. It keeps the exact words you said, the jokes, the tone, and the full context. It's there so the AI remembers how you said things.
The Event Calendar (Event Calendar): This is the magic part. Chronos reads the Storybook and pulls out only the important "moments" that happened at specific times. It turns sentences like "I bought a coffee machine on Tuesday" into a structured entry: Event: Bought Coffee Machine | Date: [Specific Date].

The Analogy: Imagine a detective's case file. The Storybook is the full transcript of the witness interview (long, detailed, maybe rambling). The Event Calendar is the sticky notes the detective put on the wall with just the facts: "Suspect arrived at 5 PM," "Car left at 6 PM." When you ask a question, Chronos looks at the sticky notes first to find the time, then opens the transcript to get the details.

2. The Smart Translator (Temporal Normalization)

Humans speak in vague time terms. We say "last week," "a few days ago," or "when I was on vacation." Computers hate this because they need exact dates.

Chronos has a special translator. When you say "last week," it doesn't just store the words "last week." It calculates the exact dates based on when you said it and writes them down as a specific range (e.g., "March 10–March 17"). This allows the AI to do math with time, like figuring out what happened "the week after" a specific event.

3. The Custom Detective (Dynamic Prompting)

In the past, AI assistants used the same search method for every question. If you asked, "Who is my boss?" and "What did I do last Tuesday?", the AI would search the same way.

Chronos is different. Before it even starts searching, it acts like a custom detective who reads your question and creates a unique game plan.

If you ask about a preference ("What coffee do I like?"), it tells the AI: "Look for opinions and habits."
If you ask about time ("What did I do last March?"), it tells the AI: "Ignore opinions, focus strictly on the Event Calendar and filter by dates."

It essentially writes a custom set of instructions for itself for every single question you ask.

4. The Iterative Search (Tool-Calling)

If the first search doesn't find the answer, Chronos doesn't just give up and guess. It has a loop where it can say, "Hmm, I found a mention of a trip, but I need to know the exact dates. Let me search the calendar again specifically for dates." It keeps refining its search until it is confident, rather than rushing to an answer.

Why Does This Matter?

The paper tested Chronos against other top AI memory systems. The results were impressive:

Chronos Low (using a standard smart model) got 92.6% of questions right.
Chronos High (using a super-smart model) got 95.6% right.

This is a huge jump compared to previous systems, which struggled with anything involving time or events that happened across different weeks.

The Bottom Line

Chronos solves the problem of "AI amnesia" regarding time. Instead of drowning in a sea of text, it builds a structured timeline of your life while keeping the full story intact. It allows an AI to not just remember what you said, but when you said it, and how that connects to everything else you've done over the last few months.

It's the difference between a friend who remembers your conversation vaguely and a friend who has a perfect, organized calendar of your life, ready to tell you exactly what you did last Tuesday.

1. Problem Statement

While Large Language Models (LLMs) have advanced conversational AI, existing memory systems struggle with long-term, temporally grounded reasoning. Current approaches face a fundamental trade-off:

Comprehensive Knowledge Graphs: Systems that extract all facts and relationships at ingestion time create large, complex structures. This introduces "context entropy" (irrelevant information) and high overhead, often failing to isolate specific temporal queries.
Pure Turn-Level Retrieval: Systems that search raw conversation turns lack structured temporal representations. They struggle with time-sensitive queries requiring date calculations, relative time reasoning (e.g., "the week after my vacation"), or cross-session event aggregation.
The Gap: No existing approach achieves query-conditioned selective extraction. There is a need for a system that structures only the temporal information relevant to specific questions while preserving raw dialogue for semantic understanding, avoiding the overhead of global extraction.

2. Methodology: The Chronos Architecture

Chronos is a conversational memory framework designed to balance structured temporal reasoning with semantic context preservation. It utilizes a Dual-Indexing strategy and Dynamic Prompting.

A. Dual-Indexing Calendars

Chronos maintains two complementary indexes:

Event Calendar: Contains extracted, structured event tuples.
- Extraction: Uses an LLM to identify events with a <subject, verb, object> structure.
- Temporal Normalization: Converts natural language time references (e.g., "last month," "recently") into precise ISO 8601 datetime ranges (start and end times). This handles ambiguity by creating windows rather than point estimates.
- Lexical Aliases: Generates 2–4 paraphrased aliases for each event using different vocabulary to improve keyword search recall (e.g., "bought Fitbit" $\rightarrow$ "picked up a fitness tracker").
- Indexing: Events are embedded and stored separately from raw text.
Turn Calendar: Preserves the raw conversational exchanges (turns) with their original context. This ensures full semantic understanding and continuity that structured events might miss.

B. Dynamic Prompting

Instead of static prompts, Chronos employs query-conditioned retrieval guidance:

Process: Upon receiving a user query, a meta-prompt analyzes the question structure to extract targets (entities, temporal constraints, operations).
Output: It generates a custom instruction preamble (1–5 bullets) directing the agent on what to retrieve, how to filter by time, and how to approach multi-hop reasoning.
Benefit: This adapts the retrieval strategy dynamically for different query types (e.g., temporal filtering vs. preference recall) without requiring separate classifiers.

C. Retrieval and Reasoning Pipeline

The system operates in a multi-stage pipeline:

Initial Retrieval: Performs dense vector search on the Turn Calendar to retrieve top candidates, followed by cross-encoder reranking (Cohere Rerank v3) and context expansion (adding adjacent turns).
Agent Reasoning (ReAct Pattern): The Chronos Agent uses a tool-calling loop to iteratively refine answers.
- Tools: Access to vector search and grep (exact keyword matching) for both the Event and Turn calendars.
- Logic: The agent decides whether to answer directly, invoke search tools for missing data, or use grep for specific entity matching. It can constrain searches by datetime ranges and cross-reference events with source dialogue.

3. Key Contributions

Query-Conditioned Selective Extraction: Chronos is the first architecture to extract only temporally grounded events for structured indexing while retaining raw dialogue for semantic retrieval, avoiding the overhead of comprehensive fact extraction.
Multi-Resolution Temporal Normalization: A novel method for converting ambiguous natural language time references into executable datetime ranges, enabling precise filtering and relative date calculations.
Dynamic Prompting for Memory: Extends query rewriting concepts to long-term memory by generating tailored retrieval guidance per question, allowing the agent to adapt its search strategy dynamically.
Dual-Index Hybrid Retrieval: Combines the simplicity of turn-level retrieval with the precision of structured event calendars, enabling both semantic understanding and rigorous temporal reasoning.

4. Experimental Results

The system was evaluated on the LongMemEvalS benchmark, comprising 500 questions across six categories (Knowledge Update, Multi-Session Aggregation, Single-Session Recall, Temporal Reasoning, etc.).

Performance:
- Chronos Low (GPT-4o): Achieved 92.60% accuracy, setting a new state-of-the-art (SOTA) with a 7.67% relative improvement over the best prior system (EmergenceMem Internal).
- Chronos High (Claude Opus 4.6): Achieved 95.60% accuracy, the highest reported score on this benchmark (+3.02% over prior records).
Category Breakdown:
- Significant gains were observed in Multi-Session Aggregation (91.73%) and Temporal Reasoning (90.23%), areas where prior methods struggled.
- Perfect accuracy (100%) was achieved on Single-Session Assistant Recall and Knowledge Update tracking.
Ablation Studies:
- Removing the Event Calendar caused the steepest performance drop (accuracy fell by ~34.5% for Chronos Low), proving the critical role of structured temporal events.
- Removing Dynamic Prompting or Initial Retrieval caused drops of 14–16%, highlighting the importance of guided retrieval.
- The Event Calendar alone accounted for a 58.9% gain over the baseline.

5. Significance

Redefining Long-Term Memory: Chronos demonstrates that high-fidelity long-horizon memory does not require comprehensive knowledge graph construction. Structuring fine-grained temporal events while preserving full dialogue context is sufficient for accurate reasoning.
Efficiency vs. Accuracy: It solves the "context entropy" problem by only structuring what is necessary for time-based reasoning, keeping the system lightweight compared to global extraction methods.
Scalability: The architecture is model-agnostic and scales effectively, showing that even with older models (GPT-4o), the architectural innovations yield SOTA results.
Future Direction: The paper suggests that future systems should not only retrieve structured events but also learn from them (e.g., updating model weights) and improve scalability for multi-agent interactions.

In summary, Chronos bridges the gap between unstructured conversational history and structured temporal reasoning, providing a robust framework for AI agents to maintain accurate, time-aware memories over extended periods.