Find the context that matters.

Search memory with the precise workflow coding agents already understand: find a file, search its contents, read the relevant lines, and make a targeted edit.

What can my agent search?

Use memory_glob to locate notes by filename or folder pattern. Use memory_grep for exact phrases or regular expressions inside current note contents. Results include file paths, versions, line numbers, and surrounding context.

Find notes:    memory_glob(pattern="**/decisions/*.md")
Search text:   memory_grep(pattern="release checklist", path="/projects")
Search regex:  memory_grep(pattern="invoice|receipt", literal=false)
Read context: memory_read(path="/projects/context.md")

These are illustrative tool calls. Ask naturally—“Find the decision about our release process”—and let your agent select the tools. The web memory desk has a filename filter for the current folder; content and pattern search are available through the MCP tools.

Does MarkdownMem use ripgrep?

It provides familiar grep and glob capabilities through MCP, rather than invoking the ripgrep executable. Searches operate over authorized memory and decrypt content for the request. They do not materialize your store as a plaintext filesystem. This is a workflow comparison, not a claim of identical implementation, regex syntax, or benchmark performance.

Which search features are available?

• Literal matching by default, with optional case-insensitive matching.
• Time-limited, Python-compatible regular expressions for single-line matches.
• Filename patterns with *, ?, character classes, and recursive ** segments.
• A file or directory boundary plus a relative filename filter.
• Matching lines with context, or just the filenames containing a match.
• Bounded pages of results, so the agent can retrieve useful pieces instead of the whole store.

Are results always exhaustive?

No. Search has scan, time, and output limits. Agents should follow continuation fields and narrow the path or pattern when a scan limit is reported. An incomplete search with no matches does not prove a note is absent. Search covers current note contents; retained history is read through the history tools.

How does an agent change a matching note?

Read the current note before editing. memory_edit replaces exact text and requires the current version; full writes and appends also check versions. Conflicting edits are rejected so the agent can read the latest content and reconcile the change.

Is this semantic or vector search?

No. This search is explicit text and filename matching. You can see why a note matched. It does not require embeddings, and it does not promise to find a concept when the note uses unrelated wording. Good note names and a small table of contents help agents choose useful searches.

See the open-source engine for the implementation, and writing useful memory for organization tips.