r3dbars/transcripted/transcripted-search-memory

Transcripted Search Memory

Version: 0.1.0

Purpose

Find evidence across Transcripted meetings and dictations, then answer with source-backed memory. Search first, synthesize second.

Workflow

1. Restate the query as a search target: topic, person, project, phrase, decision, action, or date range.
2. Prefer Transcripted MCP tools when available. Fall back to local meeting and dictation markdown files only when MCP is unavailable.
3. Combine available retrieval modes:

• exact search
• relevant context search
• recency
• date filters
• people or speaker names
• topic/project terms

1. Show the best matches before or alongside the answer.
2. Synthesize only from the evidence found.
3. If the topic appears over time, include a short timeline.
4. If evidence is thin or missing, say so plainly.

Default Output

Use these sections when relevant:

## Answer
 
## Best Matches
 
## Timeline
 
## Related Threads
 
## Confidence
 
## Sources
 
## Uncertainty

Match Format

For each strong match, include as much of this as is available:

-Source:
-Date:
-Speaker:
-Timestamp:
-Snippet:
-Why this surfaced:

Rules

• Do not answer from vague memory when sources are available.
• Do not hide weak evidence. Use low confidence when matches are sparse, old, ambiguous, or conflicting.
• If nothing relevant is found, say not found and list what was searched.
• If the user asks for "when", prefer a timeline over a prose-only answer.
• If the user asks "where did this come from", prioritize the earliest strong source.
• Keep snippets short and cite the original capture instead of pasting large transcript chunks.

Quality Check

Before answering, check:

• The answer has at least one source unless explicitly reporting no results.
• Important claims point to filenames, dates, speakers, or timestamps when available.
• Conflicting or evolving decisions are described as a timeline, not flattened into one claim.
• Search scope is clear enough that the user can ask a better follow-up.