Experiment: indexing the agent memory repo into codebase-memory-mcp

Question (task #75 / the “free experiment”): can codebase-memory-mcp give my markdown memory repo (~/.claude/projects/-home-agent-zara/memory/, 122 files, 676K, git repo, 325 [[wikilinks]]) a useful semantic knowledge-graph / better recall than the current MEMORY.md-index + grep?

What I did

index_repository(repo_path=<memory dir>, mode=full) → indexed: 247 nodes, 246 edges. Then search_code("checklist", …, mode=files) and get_architecture(aspects=["all"]).

Findings (decisive)

  • search_code works as fast graph-augmented grep: “checklist” → 11 ranked memory files in 4ms, 2.7x dedup. Genuinely handy for “which memories mention X” across all 122.
  • BUT the graph is trivial for prose. Node labels: File(122), Module(122), Section(2), Project(1). Edge types: ONLY DEFINES(124) + CONTAINS_FILE(122) — pure structural scaffolding, no semantic nodes.
  • The 325 [[wikilinks]] — the ACTUAL latent knowledge graph — were NOT captured as edges. The indexer resolves code references (LSP call/import graph); it does not parse markdown wikilinks, so the richest signal in my memory is ignored.
  • No semantic clustering of related learnings; search_code is grep, not embedding similarity, for prose.

Verdict

codebase-memory-mcp on the memory repo = fast file-search only (marginal over MEMORY.md + grep -r). It does NOT deliver the semantic/wikilink knowledge-graph the memory-layer goal wants. It’s the wrong tool for prose memory — it’s built for code call-graphs. Recommendation for #75: don’t adopt codebase-memory-mcp as the memory layer. If we want a real memory knowledge-graph, the promising directions are: (a) a purpose-built index that parses the [[wikilinks]] into edges + embeds each memory body for semantic recall (this is essentially the local-skills-embeddings-db design already scoped — [[local-skills-embeddings-db-project]]); or (b) Obsidian (native wikilink graph) / Mem0 (vector+graph). The wikilink graph is the asset to exploit; a code indexer throws it away.

Cleanup note

The index persists as project home-agent-zara-.claude-projects-home-agent-zara-memory in codebase-memory-mcp. Harmless (local, small). Can delete_project it if we want a clean slate; leaving it gives free fast search.