AuthorsMartin Vogel, Falk Meyer-Eschenbach, Severin Kohler et al.
TL;DR
Codebase-Memory uses a Tree-Sitter-based knowledge graph in SQLite exposed via MCP tools to answer code-structure questions with 10× fewer tokens at 83% quality vs. 92% for a file-exploration agent.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM coding agents repeatedly read files and grep search, consuming hundreds of thousands of tokens per task without structural understanding.
This makes structural questions like impact analysis expensive and slow, because agents lack call graphs, dependency chains, and module boundaries in a queryable form.
HOW IT WORKS
Codebase-Memory centers on a three-stage pipeline with Parse stage, Build stage, Serve stage, plus a FunctionRegistry and Louvain communities stored in SQLite and exposed via the MCP tool interface.
You can think of Codebase-Memory as turning a codebase into a card catalog: files become graph nodes, relationships become edges, and MCP tools are the librarian answering structural questions.
This design lets Codebase-Memory answer structural queries like impact analysis or hub detection directly from the graph, instead of re-reading raw files into a limited context window.
DIAGRAM
This diagram shows how Codebase-Memory runs the six pipeline phases to turn source files into a Tree-Sitter-based knowledge graph.
DIAGRAM
This diagram shows how Codebase-Memory is evaluated by comparing an MCP Agent using MCP tools against an Explorer Agent using file reading and grep.
PROCESS
Parse stage
Codebase-Memory walks Tree-Sitter ASTs across 66 languages, extracting definitions, call sites, imports, and traits into the Parse stage.
Build stage
Codebase-Memory runs the multi-pass pipeline in the Build stage, using parallel worker pools and the FunctionRegistry to assemble nodes and edges.
Serve stage
Codebase-Memory flushes the graph into SQLite, computes Louvain communities, and exposes everything via the MCP tool interface in the Serve stage.
Incremental synchronization
Codebase-Memory watches files, uses XXH3 hashes to detect changes, and re-runs the relevant pipeline phases to keep the knowledge graph fresh.
KEY CONTRIBUTIONS
Knowledge-graph architecture for code
Codebase-Memory combines Tree-Sitter parsing across 66 languages, a multi-phase Build stage, FunctionRegistry, and Louvain communities into a single SQLite knowledge graph with zero external dependencies.
MCP-based structural tool interface
Codebase-Memory exposes 14 typed MCP tools, including search_graph, trace_call_path, query_graph, and get_architecture, enabling sub-millisecond structural queries for any MCP-compatible agent.
Head-to-head evaluation across 31 languages
Codebase-Memory achieves 0.83 answer quality versus 0.92 for a file-exploration agent, with ten times fewer tokens and 2.1 times fewer tool calls on 31 real-world repositories.
RESULTS
Quality score
0.83
-0.09 vs Explorer Agent quality 0.92
Tool calls per question
2.3
-2.5 vs Explorer Agent 4.8 tool calls
Tokens per question
1000
-9000 vs Explorer Agent ∼10,000 tokens
Query latency
1
<1 ms vs Explorer Agent 10–30 s latency
On a benchmark of 12 question categories across 31 languages, Codebase-Memory delivers 0.83 quality versus 0.92 for the Explorer Agent while cutting tokens and tool calls dramatically. This shows that Codebase-Memory can trade a small quality gap for 10× token savings and >100× faster structural queries.
BENCHMARK
On a benchmark of 12 question categories across 31 languages, Codebase-Memory delivers 0.83 quality versus 0.92 for the Explorer Agent while cutting tokens and tool calls dramatically. This shows that Codebase-Memory can trade a small quality gap for 10× token savings and >100× faster structural queries.
BENCHMARK
Average quality score for structural code questions across 31 languages.
KEY INSIGHT
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.
This is surprising because many assume structural indexing would require more complexity and overhead, yet Codebase-Memory actually simplifies agent behavior and still matches or exceeds the Explorer Agent on 19 of 31 languages for graph-native queries.
WHY IT MATTERS
Codebase-Memory makes structural code retrieval a first-class capability, letting agents query call graphs, communities, and impact analysis directly via MCP tools.
This enables builders to create coding agents that start from a rich structural map instead of blind file crawling, making large, polyglot repositories and even Linux-kernel-scale projects practical to explore interactively.
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.
Andy Nguyen, Danh Doan et al.
· 2026
ByteRover organizes memory through an LLM-driven Agent Layer, a sequential Execution Layer, and a file-based Context Tree with an Adaptive Knowledge Lifecycle. On LoCoMo, ByteRover reaches 96.1% overall accuracy versus 89.9% for HonCho, and on LongMemEval-S ByteRover scores 92.8% overall.
Answers use this explainer on Memory Papers.
Checking…