memory-search

Memory Search

Search across indexed Plaud transcripts using vector similarity (Qdrant), knowledge graph (Neo4j), and keyword matching. Unlike find_recordings which searches metadata, memory search searches inside transcript content.

Tools

memory_search

Main search tool across indexed transcripts.

Parameter	Type	Default	Description
query	string	required	Natural language search query
search_type	string	"smart"	Search strategy (see table below)
since	string	null	Relative ("7d") or absolute ("2026-01-15")
until	string	null	Same format as since
last	string	null	Shorthand temporal filter ("7d", "2w", "1m")
speaker	string	null	Filter by speaker name
source	"call" | "meeting" | "desktop"	null	Filter by recording type
file_ref	string	null	Restrict to one recording. Required for "related" type
limit	int	10	Max results

Search Types

search_type	Method	Best for
smart	Hybrid (keyword + vector)	General questions -- start here
keyword	Exact matching	Specific terms, names, acronyms
semantic	Vector similarity	Conceptual queries ("discussions about scaling")
deep	Chain-of-thought reasoning	Complex, multi-hop questions
graph	Neo4j entity graph	Relationships, connections between people/topics
related	Similar content	"More like this" -- requires file_ref
entities	Entity extraction	Who, what, decisions, action items

Search Strategy (Simple to Thorough)

1. Start with smart -- hybrid search covers most queries
2. Try keyword -- if searching for exact names, terms, or acronyms
3. Try deep -- if smart results are insufficient for complex questions
4. Use graph -- for relationship questions ("who talked to whom about X")
5. Use entities -- for summaries of speakers, decisions, action items
6. Use related -- to find recordings similar to a specific one

Examples

General topic search:

memory_search(query="product roadmap discussion")

Search by speaker in recent recordings:

memory_search(query="budget", speaker="Alice", last="30d")

Find exact terms:

memory_search(query="Series B", search_type="keyword")

Complex multi-hop question:

memory_search(query="what decisions were made about the API redesign and who owns the follow-ups", search_type="deep")

Entity extraction:

memory_search(query="Q1 planning meetings", search_type="entities")

Find relationships:

memory_search(query="Alice", search_type="graph")

Find similar recordings:

memory_search(query="similar content", search_type="related", file_ref="4f75")

Prerequisites

Recordings must be indexed via memory_ingest before they appear in search results. Use get_recording(file_ref) to check indexed status. If a recording is not indexed, use the bulk-operations skill to ingest it.

Anti-Patterns (NEVER DO)

• Never assume zero results means nothing exists -- try alternative queries, different search types, or broaden terms
• Never use only keyword for conceptual queries -- use smart or semantic instead
• Never use related without file_ref -- it will error
• Never skip the prerequisites check -- unindexed recordings return no results silently
• Never present unsorted results -- sort by relevance or date (most recent first)