AuthorsAndy Nguyen, Danh Doan, Hoang Pham et al.
TL;DR
ByteRover uses an LLM-curated hierarchical Context Tree with a 5-tier retrieval pipeline to reach 96.1% accuracy on LoCoMo, +6.2 points over HonCho.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing Memory-Augmented Generation systems treat memory as an external service, so the system that stores knowledge does not understand it.
This separation leads to semantic drift, lost coordination context, and recovery fragility, where autonomous agents misremember tasks and cannot reliably reconstruct long-horizon state.
HOW IT WORKS
ByteRover centers on an Agent Layer, Execution Layer, and Context Tree connected to MiniSearch and a query cache on the local filesystem.
You can think of ByteRover like a programmer who both writes and organizes code, instead of dumping snippets into a blind search index.
This design lets ByteRover maintain structured, evolving memory with explicit relations and lifecycle metadata that a plain context window or vector store cannot provide.
DIAGRAM
This diagram shows how ByteRover routes a query through cache checks, MiniSearch, and escalating LLM reasoning tiers.
DIAGRAM
This diagram shows how ByteRover processes benchmark conversations and is scored with LLM-as-a-Judge against baselines.
PROCESS
5-Tier Progressive Retrieval
ByteRover receives a query and computes a fingerprint while checking the query cache and fuzzy cache before touching the Context Tree or MiniSearch.
MiniSearch full text search engine
If caches miss, ByteRover uses MiniSearch with BM25, fuzzy matching, and field boosts to rank Context Tree entries and detect out-of-domain queries.
Optimized LLM call
For medium confidence scores, ByteRover prefetches top documents and issues a constrained LLM call with a 1,024 token budget and temperature 0.3.
Full agentic loop
For hard or novel queries, ByteRover runs a multi turn agentic loop with tool calls over the Context Tree, up to 50 iterations and 2,048 tokens.
KEY CONTRIBUTIONS
Agent-native memory architecture
ByteRover makes curate and search tools first class in the Agent Layer, eliminating external pipelines and enabling a stateful feedback loop for memory operations.
Context Tree with Adaptive Knowledge Lifecycle
ByteRover introduces the Context Tree with importance scores in [0,100], maturity tiers from draft to core, and recency decay with τ = 30 days.
5-tier progressive retrieval strategy
ByteRover designs a 5 tier retrieval pipeline that resolves most queries at sub 100 ms without LLM calls and signals out of domain queries explicitly.
RESULTS
Overall accuracy LoCoMo
96.1%
+6.2 over HonCho
Multi-Hop LoCoMo
93.3%
+9.3 over HonCho
Overall accuracy LongMemEval-S
92.8%
+0.2 over Chronos† low backbone
Cold query latency p50
1.6 s
LongMemEval-S end to end without external infrastructure
On LoCoMo, which stresses long range conversational memory, ByteRover reaches 96.1% overall accuracy versus 89.9% for HonCho. On LongMemEval-S, which spans 500 questions and 23,867 documents, ByteRover achieves 92.8% overall accuracy while maintaining low cold query latency.
BENCHMARK
On LoCoMo, which stresses long range conversational memory, ByteRover reaches 96.1% overall accuracy versus 89.9% for HonCho. On LongMemEval-S, which spans 500 questions and 23,867 documents, ByteRover achieves 92.8% overall accuracy while maintaining low cold query latency.
BENCHMARK
Overall accuracy (%) on LoCoMo across memory systems.
BENCHMARK
Overall accuracy (%) on LongMemEval-S across selected systems.
KEY INSIGHT
Routing all queries to the full agentic loop drops ByteRover from 92.8% to 63.4% on LongMemEval-S, a 29.4 point loss.
This is surprising because many developers assume more agentic reasoning always helps, but ByteRover shows structured tiered retrieval is far more reliable.
WHY IT MATTERS
ByteRover unlocks agent native memory where the same LLM curates, structures, and retrieves long term knowledge without external databases.
Builders can now ship autonomous agents with durable, inspectable markdown memories, strong benchmarks, and zero vector or graph infrastructure to maintain.
Pratyay Banerjee, Masud Moshtaghi et al.
· 2026
APEX-MEM constructs a semi-structured temporal property graph using Ontology, Entity and Property Resolution, Fact Extraction, and Graph Agents to store and query conversational memory. On LOCOMO, APEX-MEM with GPT5 achieves 88.88% overall accuracy, beating MIRIX at 85.38% by 3.50 percentage points.
Yasong Fan
· 2026
Fan Duality Model (FDM) uses the Fan Operator, Local-Global Cache, Freeze-Scan Training, and Holographic Reference Beam Decoding to separate wave-like compression from particle-like associative recall. On WikiText-103, Fan Duality Model (FDM) reaches 64.9 perplexity with Freeze-Scan and 62.79 with holographic decoding, while achieving 0.966 MQAR accuracy compared to Transformer at 0.606.
Martin Vogel, Falk Meyer-Eschenbach et al.
· 2026
Codebase-Memory parses repositories with a multi-pass pipeline using the Parse stage, Build stage, Serve stage, FunctionRegistry, Louvain communities, and MCP tool interface to build a persistent SQLite knowledge graph. On a 31-language benchmark, Codebase-Memory reaches 0.83 quality versus 0.92 for an Explorer Agent while using ten times fewer tokens and 2.1 times fewer tool calls.
Answers use this explainer on Memory Papers.
Checking…