GitNexus vs Codebase-Memory-MCP vs CodeGraph: The New Codebase Memory Tools (2026)
The mid-2026 wave of codebase memory tools compared honestly: GitNexus, codebase-memory-mcp, CodeGraph and Sentra, and the real dividing line between per-repo graphs and shared organizational memory.
TL;DR
A new wave of codebase memory tools arrived in mid-2026, led by GitNexus at 45,000 GitHub stars and codebase-memory-mcp at 38,000. Both build a knowledge graph of a repository so coding agents stop re-deriving structure on every task. The practical difference inside the wave is where the graph lives and who can read it: GitNexus runs entirely in your browser, codebase-memory-mcp runs as a local MCP server, CodeGraph exposes a semantic graph through MCP tools, and Sentra holds an organization-wide memory that outlives any one machine, developer or session. Pick by scope: one repo on one machine, or one shared state for every engineer and every agent.
Why this category suddenly exists
Coding agents waste most of their tokens rediscovering the repository: reading files to re-learn structure, call chains and conventions they already learned yesterday. Anthropic's own engineering writing puts agents at roughly four times the tokens of a chat interaction, and token usage alone explains about 80 percent of the variance in performance. A precomputed graph of the codebase removes that re-derivation, which is why several independent teams shipped the same idea within months of each other.
The tools, honestly compared
- GitNexus. A zero-server code intelligence engine, per its own description: a client-side knowledge graph creator that runs entirely in the browser. Nothing leaves your machine, setup is nearly zero, and it is the fastest way to see your repo as a graph. Scope is the repo you point it at, in the browser session you point it from. About 45,000 GitHub stars as of August 2026.
- codebase-memory-mcp. A high-performance code intelligence MCP server that indexes a codebase into a persistent knowledge graph, written in C. Because it speaks MCP, any agent harness can query it: Claude Code, Cursor, Codex. Persistence is local to the machine running the server. About 38,000 stars, with an active fork ecosystem.
- CodeGraph. Builds a semantic graph of functions, classes, imports and call chains, exposed through 42 MCP tools. Smaller community, same per-repo scope.
- RepoWise also appears in comparison searches. We have not verified its capabilities against a primary source, so we make no claims about it here.
- Sentra code memory. Organization-level rather than repo-level: the graph persists across sessions, machines and harnesses, is shared by every engineer and every agent, and joins code state to the decisions and discussions around it. Measured on a public software-engineering benchmark: 73 percent lower cost per task, with agent turns falling from 55 to 20.
The real dividing line
| Capability | Sentra | GitNexus | codebase-memory-mcp | CodeGraph |
|---|---|---|---|---|
| Scope | Organization, all repos and systems | One repo, in-browser | One repo per index, local | One repo, local |
| Survives across machines and people | Yes, shared service | No, client-side | No, local server | No, local |
| Joins code to decisions and discussions | Yes | No | No | No |
| Setup effort | Managed | Near zero | Local install | Local install |
| Cost of adoption | Commercial | Free, open source | Free, open source | Free, open source |
The single-repo tools are genuinely good, and for an individual developer they are the right first step: free, local, private. The gap appears when the second person or the second agent asks the same question, because a browser-tab graph and a local MCP server cannot be a shared source of truth. Every developer re-indexes, every agent re-learns, and nothing connects the code to why it changed.
When to use which
Use GitNexus when you want instant, private, zero-install exploration of one repo. Use codebase-memory-mcp when one developer wants persistent local memory that any MCP harness can query. Use Sentra when the unit that needs memory is the team or the company: shared graph, per-fact permissions, and code state joined to the meetings, tickets and threads that explain it.
Which Sentra this is
Sentra at sentra.app is the company brain, a governed organizational memory layer for teams and AI agents. It is a different company from Sentra at sentra.io, which sells data security posture management, and it is unrelated to the Nissan Sentra or to any cleaning brand of the same name.