ByteRover: Agent-Native Memory Through LLM-Curated Hierarchical Context

Andy Nguyen; Danh Doan; Hoang Pham; Bao Ha; Dat Pham; Linh Nguyen; Hieu Nguyen; Thien Nguyen

cs.AIApr 2, 2026

ByteRover: Agent-Native Memory Through LLM-Curated Hierarchical Context

Authors: Andy Nguyen, Danh Doan, Hoang Pham, Bao Ha, Dat Pham, Linh Nguyen, Hieu Nguyen, Thien Nguyen,

Organizations: ByteRover

Abstract

Memory-Augmented Generation (MAG) extends large language models with external memory to support long-context reasoning, but existing approaches universally treat memory as an external service that agents call into, delegating storage to separate pipelines of chunking, embedding, and graph extraction. This architectural separation means the system that stores knowledge does not understand it, leading to semantic drift between what the agent intended to remember and what the pipeline actually captured, loss of coordination context across agents, and fragile recovery after failures. In this paper, we propose ByteRover, an agent-native memory architecture that inverts the memory pipeline: the same LLM that reasons about a task also curates, structures, and retrieves knowledge. ByteRover represents knowledge in a hierarchical Context Tree, a file-based knowledge graph organized as Domain, Topic, Subtopic, and Entry, where each entry carries explicit relations, provenance, and an Adaptive Knowledge Lifecycle (AKL) with importance scoring, maturity tiers, and recency decay. Retrieval uses a 5-tier progressive strategy that resolves most queries at sub-100 ms latency without LLM calls, escalating to agentic reasoning only for novel questions. Experiments on LoCoMo and LongMemEval demonstrate that ByteRover achieves state-of-the-art accuracy on LoCoMo and competitive results on LongMemEval while requiring zero external infrastructure, no vector database, no graph database, no embedding service, with all knowledge stored as human-readable markdown files on the local filesystem.

Figures & tables

Panel of Figure 1 — Figure 1: Architectural overview of ByteRover . Clients (TUI, CLI, MCP) connect via Socket.IO to a daemon that manages a per-project task queue and agent pool. Each agent process contains three logical layers: (1) an Agent Layer where curate and search_knowledge are first-class tools in the LLM’s reasoning loop; (2) an Execution Layer with a query executor for 5-tier progressive retrieval and a sandboxed curation environment; and (3) a Knowledge Layer with the Context Tree , BM25 full-text index, and query cache, all backed by the local filesystem with no external infrastructure.

Operation	Behavior
ADD	Create new entry; auto-generate context.md at each hierarchy level
UPDATE	Replace content of an existing entry
UPSERT	Add if new, update if exists (reduces pre-check overhead)
MERGE	Combine two entries intelligently; delete the source
DELETE	Remove a single entry or an entire subtree

Table 1: The five atomic curate operations. Every operation carries a reason field that serves as an audit trail.

Panel of Figure 2 — Figure 2: The 5-tier progressive retrieval pipeline. Search is initiated in parallel with fingerprint computation. Tiers 0–1 resolve from cache without awaiting search. Tier 2 serves high-confidence results directly from MiniSearch. Only novel or ambiguous queries escalate to Tier 3 (single optimized LLM call with pre-fetched context) or Tier 4 (full agentic loop with tool access). Approximate latencies shown on right.

Tier	Mechanism	Latency	Condition
0	Exact cache hit	$\sim$ 0 ms	Hash match + valid fingerprint
1	Fuzzy cache (Jaccard)	$\sim$ 50 ms	Jaccard $\geq\theta_{\text{fuzzy}}$
2	Direct MiniSearch	$\sim$ 100 ms	BM25 score $\geq\theta_{\text{high}}$ , sufficient gap
3	Optimized LLM call	$<$ 5 s	BM25 score $\geq\theta_{\text{med}}$
4	Full agentic loop	8–15 s	All other queries

Table 2: The five retrieval tiers with their latency characteristics and escalation conditions.

Method	Single-Hop	Multi-Hop	Open-Domain	Temporal	Overall
HonCho	93.2	84.0	77.1	88.2	89.9
Hindsight	86.2	70.8	95.1	83.8	89.6
Memobase	70.9	46.9	77.2	85.1	75.8
Zep	74.1	66.0	67.7	79.8	75.1
Mem0	67.1	51.2	72.9	55.5	66.9
OpenAI Memory	63.8	42.9	62.3	21.7	52.9

Table 3: LLM-as-Judge accuracy (%) on LoCoMo (4 categories, 1,982 questions). All systems evaluated under identical conditions using our benchmark harness with Gemini 3 Flash as the judge.

Method	KU	SSU	SSA	SSP	TR	MS	Overall
Chronos †	96.2	94.3	100.0	80.0	90.2	91.7	92.6
Hindsight	94.9	97.1	96.4	80.0	91.0	87.2	91.4
HonCho	94.9	94.3	96.4	90.0	88.7	85.0	90.4
SmartSearch †	93.6	100.0	85.7	96.7	82.7	84.2	88.4
Memora †	97.4	98.6	78.6	83.3	89.5	78.2	87.4
TiMem †	87.7	96.3	85.7	55.3	73.4	72.8	79.0

Table 4: LLM-as-Judge accuracy (%) on LongMemEval-S (500 questions). KU = Knowledge Update, SSU = Single-Session User, SSA = Single-Session Assistant, SSP = Single-Session Preference, TR = Temporal Reasoning, MS = Multi-Session. Hindsight and HonCho evaluated with our harness (Gemini 3 Flash judge). † Results from respective papers; backbone and judge configurations differ.

Metric	LoCoMo	LongMemEval-S
Context tree size	272 docs	23,867 docs
Total queries	1,982	500
Cold query latency p50	1.2 s	1.6 s
Cold query latency p95	1.4 s	2.3 s
Cold query latency p99	1.7 s	2.5 s

Table 5: Operational latency profile. Cold query latency measures end-to-end time from process invocation through response delivery, excluding answer justification and evaluation.

Configuration	KU	SSU	SSA	SSP	TR	MS	Overall	$\Delta$
w/o Tiered Retrieval	69.2	71.4	76.8	83.3	61.7	47.4	63.4	$-$ 29.4
w/o OOD Detection	98.7	98.6	98.2	96.7	89.5	85.0	92.4	$-$ 0.4
w/o Relation Graph	98.7	98.6	98.2	96.7	89.5	85.0	92.4	$-$ 0.4
ByteRover (Full)	98.7	98.6	98.2	96.7	91.7	84.2	92.8	—

Table 6: Ablation study on LongMemEval-S (500 questions). Each row disables one query-time component while keeping the curated Context Tree unchanged. Category abbreviations follow Table 4 .

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Module	Parameter	Value
Search Index	MiniSearch version	v7
	Fields	title ( $5\times$ ), content ( $1\times$ ), path ( $1.5\times$ )
	Max retrieval results	32
	Max content length	8,000 chars
Search Config	Fuzzy ratio	0.2
	Prefix matching	enabled

Appendix

Table 7: Hyperparameter configuration for ByteRover and evaluation pipeline.

Panel of Figure 3 — Figure 3: A complete knowledge entry in the Context Tree , showing the YAML frontmatter with lifecycle metadata, explicit relation annotations, raw concept (provenance), and narrative (interpreted structure).

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 373 papers

May 2, 2026

MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents

Large Language Models (LLMs) lack persistent memory for long-term personalized conversations. Existing graph-based memory systems suffer from information dilution, absent provenance tracking, and uniform retrieval that ignores query context. We introduce MemORAI (Memory Organization and Retrieval via Adaptive Graph Intelligence), a framework that integrates three innovations: selective memory filtering with dual-layer compression to retain user-persona-relevant content, a provenance-enriched multi-relational graph tracking factual origins at the turn level, and query-adaptive subgraph retrieval with Dynamic Weighted PageRank that applies query-conditioned edge weighting. Evaluated on LOCOMO and LongMemEval benchmarks, MemORAI achieves state-of-the-art performance in memory retrieval and personalized response generation, demonstrating that selective storage, enriched representation, and adaptive retrieval are essential for coherent, personalized LLM agents.

Hung Pham Van, Nguyen Manh Hieu, Khang Pham Tran Tuan +4

1Independent Researcher · 2Hanoi University of Science and Technology · 3VNU University of Engineering and Technology +1

Conversational Memory Graph Retrieval-Augmented Generation Persistent Memory for Language Models Graph-Based Agent Memory Long-Term Conversational Memory Agent Memory Management

Ask PDF

May 6, 2026

Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall

Extraction at ingestion is the wrong primitive for agent memory: content discarded before the query is known cannot be recovered at retrieval time. We propose True Memory, a six-layer architecture that shifts the center of the system from a storage schema to a multi-stage retrieval pipeline operating over events preserved verbatim. The full system runs as a single SQLite file on commodity CPU with no external database, vector index, graph store, or GPU. On LoCoMo (1,540 questions across 10 multi-session conversations), True Memory Pro reaches 93.0% accuracy (3-run mean) against 61.4% for Mem0, 65.4% for Supermemory, approximately 71% for Zep, and 94.5% for EverMemOS under a matched gpt-4.1-mini answer model. On LongMemEval (500 questions), True Memory Pro reaches 87.8% (3-run mean). On BEAM-1M (700 questions at the 1-million-token scale), True Memory Pro reaches 76.6% (3-run mean), above the prior published result of 73.9% for Hindsight. A 56-configuration ablation shows a 1.3-percentage-point spread within the top-performing configuration family.

Joshua Adler, Guy Zehavi

Sauron Labs

Agent Memory LLM Agent Memory Persistent Memory for Language Models Long-Term Conversational Memory Agent Memory Management

Ask PDF

Aug 13, 2026

RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory

LLM-based agents increasingly rely on external memory to support long-horizon reasoning and interaction. However, the main bottleneck is not simply storing past experience, but recovering the right set of evidence when relevant information is distributed across many interactions. Existing approaches struggle with this access problem. Full-context methods require noisy long-context search, flat retrieval often returns isolated and incomplete records, and graph-based memory systems can be expensive to construct while compressing rich event context. We introduce RippleMem, a long-term memory system that replaces one-shot retrieval with adaptive associative recollection. Inspired by cue-dependent episodic retrieval and associative completion, RippleMem stores interaction history as cue-rich episodic memory units and organizes them in an event-centric memory graph. Given a query, it first recalls relevant memory anchors through hybrid cues, then expands from these anchors along semantic and structural associations to recover missing supporting evidence. In this way, initially recalled memories serve not only as answer context, but also as cues for completing the evidence needed to answer. Experiments on LoCoMo and LongMemEval-S show that RippleMem achieves the best overall performance across evaluated settings, improving LLM-as-a-Judge accuracy by 3.95% on LoCoMo and up to 11.87% on LongMemEval-S, while reducing graph construction cost by about 30x.

Jingbo Ji, Lingyi Li, Xilong Cheng +4

Communication University of China, Beijing, China. · Zhilian Yinghe Technology Co., Ltd., Beijing, China · State Key Laboratory of Media Convergence and Communication, Beijing, China

Agent Memory Persistent Memory for Language Models Graph-Based Agent Memory Associative Memory

Ask PDF

Abstract

LLM Agent Memory

Explore similar work

MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents

Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall

RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory