Codebase Memory MCP | Ry Walker Research

Key takeaways

A native executable builds a persistent local code graph without a separate database service or model API.
Hybrid LSP means embedded type-resolution logic, not a fleet of external language-server processes.
The authors' study found lower token use with a measurable answer-quality tradeoff.
Version 0.11.0 changes index format and memory handling, so upgrades require a fresh initial index.

FAQ

What is Codebase Memory MCP?

An MIT-licensed structural-analysis backend that lets coding agents search and traverse repository graphs over MCP or a local CLI.

Does it contain an LLM?

No. It builds the graph locally; the connected coding agent interprets questions and uses its tools. The agent's model costs and data handling remain separate.

Is Hybrid LSP a language server?

The project uses that name for embedded C implementations of language type-resolution logic. It does not launch the corresponding external language servers.

Does it guarantee lower costs and equal accuracy?

No. The authors' March 2026 study reported about tenfold lower token use but an answer-quality score of 83% versus 92% for file exploration on its selected tasks.

Executive Summary

Codebase Memory MCP, distributed from DeusData/codebase-memory-mcp, builds persistent repository graphs for coding agents. It is a structural-analysis backend, with a CLI and MCP interface, rather than a coding agent itself. Its shipped indexing and query tools make it a member of the code intelligence tools category.^[1]

Its strongest distinguishing evidence is unusually explicit about tradeoffs: the authors' March 2026 preprint reports substantially lower token consumption alongside lower average answer quality. That makes it a candidate for measured evaluation, not a basis for promising universal savings.^[2]

Attribute	Checked September 15, 2026
Project	DeusData/codebase-memory-mcp^[3]
Current release	v0.11.0, September 15, 2026^[4]
License	MIT, with copyright and permission-notice conditions^[3]
Distribution	Native executables for macOS, Linux, and Windows; fully static Linux builds are separate portable variants^[1]

Product Overview

The agent can index repositories, find symbols, follow calls, inspect architecture, map diffs to affected code, and request source snippets. The tool also supports architecture-decision records and runtime traces for validating HTTP relationships. The published tools include both read operations and mutations such as index deletion.^[1]

Surface	Role
MCP server	Structural context for a connected coding agent
One-shot CLI	Query or index from scripts without a standing client session
Local graph viewer	Interactive inspection of the indexed graph

All three are documented public capabilities, not separately priced hosted plans.^[1]

Technical Architecture

The implementation combines Tree-sitter extraction with persistent SQLite storage and additional type-resolution logic for selected languages. Its “Hybrid LSP” label describes embedded C algorithms inspired by language-server behavior; it does not mean those external servers run alongside the indexer. Broader parser coverage should not be confused with equally deep semantic resolution across languages.^[1]

Version 0.11.0 changes file-node identity to retain extensions, avoiding collisions such as auth.py and auth.ts. The first run rebuilds older indexes while preserving architecture-decision records. It also introduces compact default responses with continuation cursors and memory-budget handling that spills intermediate data or fails while preserving the prior index.^[4]

Evidence and Limitations

The March 28, 2026 preprint evaluates 31 repositories/languages with 12 standardized question categories, using Claude Opus 4.6 in both arms. The first author graded answers against manually derived references. It reports quality scores of 0.83 for graph-assisted exploration and 0.92 for file exploration, with approximately tenfold lower tokens and 2.1-fold fewer tool calls. Those are author-reported results from an earlier version, not an independent benchmark of v0.11.0.^[2]

The project's measurement guide separates answer quality, backend performance, and agent token consumption. It recommends fixed repository commits, matching prompts and budgets, isolated sessions, and retaining failures. This is the appropriate way to evaluate the tool on a team's own work; fast graph queries alone do not establish accurate answers.^[5]

Strengths

Local structural backend: no separate database service or built-in model is required for indexing and graph queries.^[1]
Explicit measurement guidance: the project explains how to compare quality and cost without confusing backend counters with model usage.^[5]
Operational improvements: the current release addresses memory, installer, daemon, and index-format reliability.^[4]

Cautions

Accuracy tradeoff: the published study did not match file exploration on overall answer quality.^[2]
Memory is a real constraint: v0.11.0 improves budget enforcement, but a graph larger than the available budget can still fail to index.^[4]
Documentation conflict: the README says background update checks have been removed, while SECURITY.md still describes one. Verify the installed release if zero network activity is a requirement.^[1]^[6]
Provenance is not a security audit: the security policy distinguishes signed build provenance from independent certification or proof that code is vulnerability-free.^[6]

What Developers Say

GitHub user shafty023 reported on September 1, 2026 that v0.10.8 exceeded configured memory budgets while indexing a private TypeScript monorepo of about 14,000 files. The report carefully distinguishes cached timing from a cold run and says the repository cannot be shared. The issue is closed, and v0.11.0 explicitly addresses its memory-budget class of failures. This is historical first-hand evidence, not a reproduction performed for this profile.^[7]^[4]

Pricing & Licensing

Component	Cost model
Local software	Free under MIT; retain required notices^[3]
Host resources	Operator supplies memory, CPU, and index storage^[4]
Connected agent	Its model and service charges remain separate; the backend contains no LLM^[1]

The security policy describes a solo, volunteer-maintained project and best-effort handling. Do not infer a support SLA from the open-source release.^[6]

Competitive Positioning

This tool is worth comparing when persistent structural queries and a native deployment fit the workflow. Its own study suggests combining graph answers with file-level investigation for questions requiring fuller context, rather than assuming either retrieval method is sufficient for every task.^[2]

Ideal Customer Profile

Best fit: developers prepared to test graph coverage, memory use, and answer quality on repeatable repository questions.

Poor fit: teams seeking a hosted support contract or treating advertised speed as proof of safe code changes.

Viability Assessment

A new release and detailed public evaluation provide evidence of current maintenance. They do not establish long-term support capacity. The immediate decision is whether the installed version produces useful, inspectable answers within the team's resource budget.^[4]^[6]

Bottom Line

Recommended for: measured trials of local structural retrieval in existing coding-agent workflows.

Not recommended for: replacing source review and tests with an assumed complete dependency graph.

Outlook: credible technical scope and candid published tradeoffs, with version-specific reliability and coverage still worth testing.

Research by Ry Walker Research • methodology

Sources