memory-t1-temporal

name	memory-t1-temporal
title	Memory-T1: RL for Temporal Reasoning in Multi-Session Agents
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.20092
keywords	["reinforcement-learning","temporal-reasoning","memory-retrieval","multi-session"]
description	Enable agents to accurately identify temporally relevant information in long multi-session dialogues through RL-based memory retrieval. Combines coarse-to-fine candidate selection with multi-level temporal consistency rewards—providing dense supervision that disambiguates time expressions and maintains coherence across 128k-token contexts.

Overview

Memory-T1 addresses a critical failure mode in long-context language models: inaccurate temporal reasoning when processing multi-session dialogues. As conversation history grows, models conflate events from different time periods. This framework adds explicit temporal reasoning via RL-trained memory retrieval.

Core Technique

The key insight is that temporal reasoning requires multi-level supervision: accuracy, evidence grounding, and temporal consistency.

Coarse-to-Fine Retrieval Strategy: Two-phase filtering manages computational complexity while preserving accuracy.

# Two-phase candidate selection for temporal reasoning
class TemporalMemoryRetrieval:
    def __init__(self):
        self.temporal_predictor = TemporalFilter()
        self.bm25_ranker = BM25Ranker()
        self.fine_grained_selector = RLPolicyModel()

    def retrieve_with_temporal_filtering(self, query, dialogue_history):
        """
        Phase 1: Coarse filtering narrows candidate pool
        Phase 2: RL-based fine selection with temporal rewards
        """
        # Phase 1a: Temporal filtering
        # Predict time range of query
        query_time_range = self.temporal_predictor(query)
        # Filter dialogue to sessions within time range
        candidate_sessions = filter_by_timerange(
            dialogue_history, query_time_range
        )

        # Phase 1b: Relevance filtering (BM25)
        relevant_candidates = self.bm25_ranker.rank(
            query, candidate_sessions, top_k=50
        )

        # Phase 2: Fine-grained RL selection
        selected_evidence = self.fine_grained_selector.select_with_policy(
            query, relevant_candidates
        )

        return selected_evidence

Multi-Level Temporal Consistency Reward: Three complementary reward signals enable dense supervision on temporal dimensions.