Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP

Martin Vogel; Falk Meyer-Eschenbach; Severin Kohler; Elias Grünewald; Felix Balzer

cs.SEMar 28, 2026

Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP

Authors: Martin Vogel, Falk Meyer-Eschenbach, Severin Kohler, Elias Grünewald, Felix Balzer

Organizations: Independent Researcher, Berlin, Germany · Institute of Medical Informatics, Charité – Universitätsmedizin Berlin, Berlin, Germany · Clinical Study Center, Berlin Institute of Health, Berlin, Germany · Institute of Informatics, Humboldt University, Berlin, Germany ·

Abstract

Large Language Model (LLM) coding agents typically explore codebases through repeated file-reading and grep-searching, consuming thousands of tokens per query without structural understanding. We present Codebase-Memory, an open-source system that constructs a persistent, Tree-Sitter-based knowledge graph via the Model Context Protocol (MCP), parsing 66 languages through a multi-phase pipeline with parallel worker pools, call-graph traversal, impact analysis, and community discovery. Evaluated across 31 real-world repositories, Codebase-Memory achieves 83% answer quality versus 92% for a file-exploration agent, at ten times fewer tokens and 2.1 times fewer tool calls. For graph-native queries such as hub detection and caller ranking, it matches or exceeds the explorer on 19 of 31 languages.

Figures & tables

Panel of Figure 1 — Figure 1: Codebase-Memory architecture. Source files are parsed via Tree-Sitter, stored as a knowledge graph in SQLite, and exposed to LLM agents through 14 typed MCP tools. A file watcher triggers incremental re-indexing on changes.

Node Type	Extracted From
Project , Package , Folder	Directory structure
File , Module	File system
Function , Method , Class	Tree-Sitter AST
Interface , Enum , Type	Tree-Sitter AST
Route	Framework detection

Table 1: Node types in the knowledge graph.

Edge Type	Semantics
CALLS , HTTP_CALLS , ASYNC_CALLS	Invocation
IMPORTS	Module import
CONTAINS_* , DEFINES , DEFINES_METHOD	Structural nesting
IMPLEMENTS	Interface impl.
HANDLES	Route $\to$ handler
INHERITS , DECORATES	OOP hierarchy

Table 2: Edge types representing code relationships.

Phase	Output
1. Structure	File discovery; Project, Package, Folder, File nodes + containment edges
2. Extraction	Parallel definition extraction via pthreads worker pool; Function, Method, Class, Interface, Enum, Type nodes; decorator tags; FunctionRegistry
3. Resolution	Parallel call, usage, and semantic resolution; CALLS, IMPORTS, USAGES, USES_TYPE, IMPLEMENTS, INHERITS, DECORATES edges
4. Enrichment	TESTS edges, HTTP route matching, config linking, git co-change edges
5. Flush	Bulk INSERT into SQLite with deferred index creation
6. Post-index	Louvain communities, XXH3 file hashes

Table 3: Pipeline phases and their outputs.

Category	Tool	Description
Indexing	index_repository	Build/update graph
	index_status	Poll indexing progress
	list_projects	List indexed repos
	delete_project	Remove index
Query	search_graph	Symbol search
	trace_call_path	Call-chain traversal

Table 4: MCP tools exposed by Codebase-Memory .

Q#	Category	Primary Tool
1	Indexing	get_graph_schema
2–3	Discovery	search_graph
4	Pattern matching	search_graph
5	Code retrieval	get_code_snippet
6	Code search	search_code
7–8	Call tracing	trace_call_path

Table 5: Benchmark question categories and primary MCP tools.

Metric	MCP	Explorer	Difference
Quality score	0.83	0.92	90% of Explorer
Tool calls / question	2.3	4.8	2.1 times fewer
Tokens / question	$\sim$ 1,000	$\sim$ 10,000	10 times fewer
Query latency	$<$ 1 ms	10–30 s	$>$ 100 times faster

Table 6: Head-to-head comparison: MCP Agent vs. Explorer Agent.

Feature	Emb./RAG	Repo-Map	Graph+LLM	Ours
Languages	10–30	$\sim$ 100	8–14	66
Struct. queries	No	No	Yes	Yes
Infra.	Vector DB	None	Neo4j	SQLite
Persistence	Yes	No	Yes	Yes
Embed. model	Yes	No	Some	No
Tok. / query	$\sim$ 2–5K	$\sim$ 1K	$\sim$ 5K	$\sim$ 1K

Table 7: Comparison of code retrieval approaches for LLM agents.

Operation	Time
Fresh index (49K nodes, 196K edges)	$\sim$ 6 s
Fresh index (Linux kernel, 2.1M nodes, 4.9M edges)	$\sim$ 3 min
Incremental re-index	$\sim$ 1.2 s
Cypher query (relationship traversal)	$<$ 1 ms
BFS call-path tracing (depth=5)	$\sim$ 0.3 ms
Name search (regex)	$<$ 10 ms

Table 8: System performance benchmarks.

AI Coding Agents

AI coding agents use language models and software tools to plan, edit, test, and maintain code with varying levels of autonomy.

See all 327 papers

LARGER: Lexically Anchored Repository Graph Exploration and Retrieval
May 8, 2026Yuntong Hu, Tongli Su, Liang Zhao +2AI Coding Agents Agentic Retrieval
Emory University
CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents
Aug 6, 2026Wuya Chen, Yihao yang, Yang Cao +1AI Coding Agents LLM Agent Workflow Optimization
1Netease Guangzhou AI Lab · 2Independent Researcher
ReCUBE: Evaluating Repository-Level Context Utilization in Code Generation
Mar 26, 2026Jiseung Hong, Benjamin G. Ascoli, Jinho D. ChoiRepository-Level Code Generation Code Generation Evaluation
Language Technologies Institute, Carnegie Mellon University · Computer Science, Emory University

May 8, 2026J/K move · Enter open · S save

LARGER: Lexically Anchored Repository Graph Exploration and Retrieval

Yuntong Hu, Tongli Su, Liang Zhao, Bowen Zhu+1

Emory University

AI Coding Agents Agentic Retrieval Graph-Based Retrieval

Open paper AskPDF arXiv

Repository-level coding agents must first localize the files and symbols relevant to a task; failures at this stage can cascade across downstream objectives ranging from patch generation to test writing and codebase question answering. Existing agents navigate repositories primarily through lexical search, often missing structural relations such as imports, call chains, type hierarchies, and code-test links. Graph-based retrieval can recover such dependencies, but existing approaches often require separate graph tools or traversal stages that fragment the agent's interaction loop. We formalize repository context localization as Lexically Anchored Structural Localization, where success depends on turning lexical matches into high-precision structural entry points and exposing the most useful confidence-filtered local neighborhoods within the agent's existing search loop. We introduce LARGER (Lexically Anchored Repository Graph Exploration and Retrieval), a lexically anchored active-set retrieval framework that starts from lexical matches, aligns them to graph anchors, and performs confidence-filtered local expansion within the agent's existing search loop. LARGER integrates directly into existing CLI coding agents without requiring external graph databases or specialized graph interfaces. Across four benchmarks spanning localization, test generation, and codebase understanding, LARGER improves file-level Acc@5 on LocBench by +13.9 points with tuned hyperparameters and still gains +11.8 points with fixed hyperparameters over the strongest baseline, while delivering consistent gains on MuLocBench, SWE-Atlas Test Writing, and SWE-Atlas Codebase QA.

Abstract

AI Coding Agents

Explore similar work

LARGER: Lexically Anchored Repository Graph Exploration and Retrieval

CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents

ReCUBE: Evaluating Repository-Level Context Utilization in Code Generation