Research Messages Search

Find past research conversations, paper discussions, and analysis from your conversation history.

Quick Start: The Research Needle

Most common use: Finding when you discussed a specific paper, topic, or research question.

Standard Opening

code

User: "When did we discuss transformer architectures?"

Agent: "Let me search our conversation history for transformer discussions.

I'll use hybrid search (semantic + keyword) to find the most relevant messages."

Then execute the search:

bash

npx tsx [SKILL_DIR]/scripts/search-messages.ts --query "transformer architectures" --limit 10

Present results:

code

Agent: "Found 8 conversations about transformers:

Most relevant:
- Dec 15, 2025: Analyzed 'Attention Is All You Need' paper
- Dec 18, 2025: Discussed transformer memory mechanisms
- Jan 3, 2026: Compared transformer vs Mamba architectures

Want me to expand around any of these to see the full context?"

Diagnosis: What Are They Looking For?

Ask 2 questions to understand the search:

Question 1: What type?

code

Agent: "Are you looking for:
A) A specific paper we discussed?
B) Analysis of a research topic?
C) Discovery results from a query?
D) Citation network we built?
E) Something else?"

Question 2: Time range?

code

Agent: "Time frame:
A) Recent (last week)
B) This month
C) Specific date/period
D) Any time (full history)"

Route based on answers:

code

Type A (paper) + Recent → Paper title search, last 7 days
Type B (topic) + Any time → Semantic search, all history
Type C (discovery) + Month → Keyword "discovery", current month
Type D (citation) + Specific → Date-bounded search

Conversation Patterns

Pattern 1: Finding Paper Discussion

Trigger: User asks about a specific paper

code

User: "What did we say about the BERT paper?"

Agent: "Let me search for BERT discussions."

[Run search with "BERT paper", focus on title/authors]

Agent: "Found 3 conversations:
- Nov 20: Initial BERT paper analysis
- Nov 22: Compared BERT vs RoBERTa  
- Dec 1: Used BERT for sentiment analysis

The Nov 20 conversation has the main analysis. Want to see it?"

[If yes, use needle+expand strategy]

Pattern 2: Finding Topic Discussions

Trigger: User asks about research topic/concept

code

User: "When did we talk about citation networks?"

Agent: "Searching for citation network discussions..."

[Run semantic search, --mode vector for concept matching]

Agent: "Found 5 conversations about citation networks:

Most relevant:
- Jan 5: Built citation graph for ML papers
- Jan 8: Analyzed influential papers using PageRank
- Jan 10: Discussed loop closure in citations

These span multiple conversations. Want the full thread?"

[If yes, expand around first occurrence and follow chronologically]

Pattern 3: Finding Discovery Results

Trigger: User asks about papers found from discovery

code

User: "What papers did we find about state space models?"

Agent: "Searching for state space model discoveries..."

[Search with "discovery state space models" + keyword mode]

Agent: "Found 2 discovery sessions:
- Dec 28: Discovered 15 SSM papers from ArXiv
- Jan 4: Found Mamba paper and variants

The Dec 28 session has the full paper list. Show that?"

Pattern 4: Needle + Expand Strategy

Trigger: User wants full context around a specific finding

code

Agent: "Found the needle! Message from Dec 15 at 2:30 PM.

Getting context around that message..."

[Run get-messages with --before and --after]

Agent: "Here's the full conversation (10 messages before + 10 after):

[Show conversation thread]

That's the complete context. Need more?"

Quick Reference Cards

Card 1: Search Modes

When to use each mode:

Mode	Use When	Example
`hybrid` (default)	General search	"transformer discussions"
`vector`	Concept/semantic	"papers about memory"
`fts`	Exact phrase	"Attention Is All You Need"

Default to hybrid unless user needs exact matches or pure semantic.

Card 2: Time Filters

Common time patterns:

bash

# Last week
--start-date "2026-01-06" --end-date "2026-01-13"

# Specific month
--start-date "2025-12-01" --end-date "2025-12-31"

# After specific date
--start-date "2025-11-01"

# Before specific date  
--end-date "2025-10-31"

Pro tip: Use ISO format YYYY-MM-DD for dates.

Card 3: Research-Specific Queries

Effective search terms for research:

•Paper titles: Use quotes for exact match
•Authors: Include "et al" or full names
•Topics: Use technical terms, not casual language
•Citations: Include "citation" or "references"
•Discovery: Include "discovered" or "found papers"
•Analysis: Include "analyzed" or "compared"

Examples:

code

"Attention Is All You Need"  → Exact paper
transformer memory mechanisms → Topic discussion
Vaswani et al                → Author search
citation network analysis    → Analysis type
discovered mamba papers      → Discovery results

Card 4: The Needle + Expand Pattern

Two-step strategy for full context:

Step 1: Find the needle

bash

npx tsx [SKILL_DIR]/scripts/search-messages.ts \
  --query "your search terms" \
  --limit 5

Step 2: Expand around it

bash

# Get context before
npx tsx [SKILL_DIR]/scripts/get-messages.ts \
  --before "message-id-from-search" \
  --limit 10

# Get context after  
npx tsx [SKILL_DIR]/scripts/get-messages.ts \
  --after "message-id-from-search" \
  --order asc \
  --limit 10

Refinement Patterns

Refinement 1: Too Many Results

Symptom: Search returns 50+ messages, overwhelming

Fix: Narrow the search

code

Agent: "That search returned too many results. Let's narrow it:

Option A: Add time filter (when roughly was this?)
Option B: Add specific terms (any unique words?)
Option C: Use exact phrase (put in quotes)

Which helps?"

Refinement 2: No Results Found

Symptom: Search returns 0 results

Fix: Broaden and diagnose

code

Agent: "No results for that search. Let's troubleshoot:

1. Try semantic search (concepts vs exact words):
   --mode vector

2. Try broader terms:
   'transformer' instead of 'transformer architecture'

3. Search all agents (if multi-agent setup):
   --all-agents

Which should we try?"

Refinement 3: Wrong Context

Symptom: Found messages but not the right conversation

Fix: Expand correctly

code

Agent: "That's not the conversation you wanted. Let me:

1. Check the other search results (we have 4 more)
2. Expand more context (20 messages instead of 10)
3. Refine search terms (what's unique about that conversation?)

Which approach?"

Refinement 4: Need Cross-Session Context

Symptom: Topic discussed across multiple sessions

Fix: Search with sort, present chronologically

code

Agent: "This topic spans multiple conversations. Here's the timeline:

Session 1 (Dec 5): Initial discussion
Session 2 (Dec 8): Follow-up analysis
Session 3 (Dec 12): Final conclusions

Want to see all three in order, or just the most relevant?"

Troubleshooting

Problem	Quick Fix
Search too slow	Reduce --limit to 5
No results found	Use --mode vector for semantic search
Too many results	Add date range with --start-date
Wrong messages	Use exact phrase in quotes
Need more context	Use get-messages.ts with --before/--after
Want full thread	Expand before+after the needle
Multi-agent search	Add --all-agents flag
Need specific date	Use --start-date and --end-date (same day)
Results not relevant	Switch to --mode fts for exact keyword

Advanced: Cross-Agent Research Search

Use case: Finding which agent discussed a topic

code

Agent: "Searching across all agents for transformer discussions..."

[Run with --all-agents flag]

Agent: "Found discussions in 3 agents:
- Agent A (Lead Engineer): Technical analysis
- Agent B (Literature Review): Paper summaries
- Agent C (Research Assistant): Discovery results

Which agent's context do you want to explore?"

Results include agent_id - use with finding-agents skill to get agent details.

Summary: The Agent's Mental Model

Core workflow:

•Understand what user is looking for (paper, topic, discovery, etc.)
•Choose search mode (hybrid default, vector for concepts, fts for exact)
•Run search with appropriate filters
•Present top results with context
•If needed, expand around the needle
•Refine if results aren't quite right

Key principles:

•Default to hybrid search (best balance)
•Use semantic (vector) for concept searches
•Use exact (fts) for specific phrases/titles
•Always offer to expand for full context
•Present results with timestamps and message types
•Help user narrow down if too many results
•Help user broaden if no results

Success: User finds the research conversation they were looking for and gets the full context they need.