cortexmd — code-nav & memory

cortexmd — code-nav & memory

cortexmd exposes MCP tools that are far cheaper than Read/Grep for TS/JS/Python/Rust/Go, plus a persistent memory + knowledge graph. Prefer them.

Cheap code-navigation (≈60 tokens/result) — prefer over Read/Grep

• code_symbol_search(query, repo) — find symbols by name/signature/docstring
• code_file_outline(repo, path) — file overview without reading the body
• code_symbol_get(id) — body of one symbol (capped ~200 lines)
• code_symbol_callers(id) / code_symbol_callees(id) — call-graph navigation
• code_change_impact(id, depth) — transitive callers ("if I change X, who breaks?")
• code_call_chain(source, target) — shortest call path
• code_find_semantic_duplicates(repo) / code_find_dead_code(repo) / code_find_import_cycles(repo)

Read/Grep are the fallback — only when the symbol is missing from the index or you need literal bytes (comments, non-source content).

Memory & knowledge graph

• memory_wakeup(agentName, preset) — boot context (preset: tiny|standard|full)
• memory_recall(query) — hybrid recall over memories + notes
• memory_store(content, ...) — store a durable fact/decision/observation (auto-links + KG seed)
• notes_upsert / notes_get / notes_search — vault notes
• kg_query / graph_neighbors / graph_traverse — knowledge-graph navigation
• memory_dream — consolidation: decay, archive, and reconcile cold notes into projects

Hooks (installed by this plugin)

• UserPromptSubmit → cortexmd recall --hook recalls relevant memory per prompt (the sweep).
• SessionStart → keeps the HUD statusline daemon alive.
• PreToolUse:Bash → rewrites grep/cat/head/tail on indexed repos to code-nav.
• PostToolUse:Bash → captures high-signal commands as memories.

These require the cortexmd binary on PATH and a running cortexmd server (see the plugin README).