Bug: plain markdown files silently dropped from RAG index (0 chunks)

[ch405canova-sudo]

Bug report

Describe the bug
Markdown files that are not named readme.md and not located under a docs/ path are silently skipped during index_documents — they produce 0 chunks even though they exist.

To reproduce

1. Create a file like knowledge/mahoo-architecture.md containing # Title\n\nsome content.
2. Run index_documents on the parent directory.
3. index_metadata shows 0 chunks for that file.

Expected behavior
Every .md file should be indexed as a docs source.

Environment

• agent-memory-mcp v0.9.1
• Ubuntu 24.04, Linux

Root cause
internal/rag/documents.go: processFile calls classifySourceType(relPath, "", "") with empty content, so the .md branch heuristic strings.Contains(contentLower, "# ") can never match. The result "" causes the file to be skipped. (There is a second call site in supportedSourceType with the same empty-content problem.)

Fix available
See PR #17 (classifies every .md file as docs, adds regression tests).

[ch405canova-sudo]

Reported and fixed jointly by chaos (@ch405canova-sudo) and opencode.

[ipiton]

Thanks for the report and the precise root-cause analysis — the diagnosis was exactly right.

This one is already fixed on main: b052f2d ("drop the heading heuristic for markdown — it hid whole knowledge bases") and bd00d87 removed the readme.md / docs/ / # conditions, so the .md branch of classifySourceType now returns docs unconditionally. The reason v0.9.1 still reproduces it is that the fix has not shipped in a release yet — git tag --contains b052f2d is empty. It will go out with the next one.

The two regression cases from #17 are worth keeping regardless, so they are picked up in #20 — the existing table only covered README.md, which is precisely why the heuristic could rot unnoticed.

Closing as already fixed on main; #18 stays open and is handled in #20.

[ipiton]

Shipped in v0.9.2 — brew upgrade agent-memory-mcp, or grab a binary from the release page. This is the release the fix was waiting for; on v0.9.1 the bug reproduces exactly as you described.

After upgrading, a re-index is needed for the previously skipped files: they were never opened, so nothing about them exists in the vector store yet.

Note that the shipped fix goes further than the heading heuristic you diagnosed — that branch is gone entirely rather than repaired. Measuring on a live repo showed only 6 of 58 memory files carried a # line, so even with correct content passed in, 52 would have stayed invisible: frontmatter-plus-prose notes carry no heading by design. Markdown reaching the classifier was already walked from a configured index dir, so its presence is the operator's intent; exclusion now lives solely in IndexExcludeDirs / IndexExcludeGlobs.

[ch405canova-sudo]

Together for a better world...