codebase-memory

---

name: codebase-memory description: Use when you need structural graph navigation — trace call paths, detect blast radius, find dead code, or query architecture — before reading files. Structural queries only: not for semantic meaning (use cocoindex-code for that).

$codebase-memory — Structural Graph Navigation

Zero-file-read structural analysis via codebase-memory-mcp. Always try these tools BEFORE reading files for structural questions.

Not for semantic search (use $cocoindex-code). Not for rename/refactor (use $serena).

Quick Decision Matrix

Question	Tool
What calls foo()?	trace_path(function_name="foo", direction="inbound")
What does foo() call?	trace_path(function_name="foo", direction="outbound")
Full call chain	trace_path(direction="both", depth=3)
Impact of my change	detect_changes(base_branch="main")
Architecture overview	get_architecture
Find by name/pattern	search_graph(name_pattern="Auth.*")
Dead code	search_graph(max_degree=0, label="Function")
Custom Cypher query	query_graph(query="MATCH (n:Function)-[:CALLS]->(m) RETURN n.name LIMIT 20")
Execution flows	list_flows → get_affected_flows

First Use

# Index the repository (one-time setup, ~6s for 500 files)
index_repository(repo_path=".")
# Verify
list_projects
# View architecture map
get_architecture

Config in .codex/config.toml

[mcp_servers.codebase-memory]
command = "uvx"
args = ["codebase-memory-mcp@latest", "--data-dir", ".codex/codebase-memory"]

Or if using the binary release directly:

[mcp_servers.codebase-memory]
command = "codebase-memory-mcp"
args = []

Workflow — Before Reviewing Changes

1. detect_changes(base_branch="main")    → risk score + blast radius
2. get_affected_flows                    → which execution paths break
3. trace_path on highest-risk functions  → full call chain context

Workflow — Codebase Exploration

1. get_architecture    → community map, entry points, god nodes
2. search_graph(name_pattern="<keyword>")  → find relevant nodes
3. trace_path          → traverse from found nodes
4. get_code_snippet(qualified_name="<name>")  → source with context

Known Gotchas

1. trace_path requires exact qualified names — use search_graph(name_pattern=...) first to find them
2. Results default to 10 per page — check has_more and paginate with offset
3. query_graph has a 200-row cap — use search_graph with degree filters for counts
4. direction="outbound" misses cross-service callers — use direction="both" to see all
5. Index is not always auto-updated — run index_repository(mode="fast") for immediate refresh

Token Efficiency

120× fewer tokens than file exploration for structural questions. A call chain that would require reading 15 files can be answered in 1-2 MCP calls.