Skip to main content

​
Overview

The memory search command finds memories using a hybrid approach: full-text search (FTS5) combined with semantic/vector search when embeddings are available.

​
Syntax

memory search <query> [OPTIONS]

​
Arguments

​
query
string
required
Search query. Can be keywords, phrases, or natural language questions.

​
Options

​
--limit
integer
default:"5"
Maximum number of results to return.
​
--project
flag
Filter results to the current project only (uses current directory name).
​
--source
string
Filter by source (e.g., cursor, claude, codex).

​
Examples

memory search "authentication"
Output: 
 Results (3 found) 

 [1] JWT expiry set to 7 days (score: 0.89)
     decision | 2026-03-01 | myproject
     What: Changed token expiration from 24h to 7d for better UX
     Why: Users were complaining about frequent re-logins
     Impact: Reduced support tickets by 40%

 [2] OAuth2 flow implemented (score: 0.76)
     pattern | 2026-02-28 | myproject
     What: Added OAuth2 authorization code flow with PKCE
     Details: available (use `memory details a8c3f9d2e1b4`)

 [3] Fixed auth middleware bug (score: 0.64)
     bug | 2026-02-25 | myproject | cursor
     What: Added null check before accessing user.role

​
Search in current project only

memory search "database" --project

​
Search with more results

memory search "performance optimization" --limit 10

​
Search by source

memory search "refactoring" --source claude

​
Natural language query

memory search "why did we choose postgres over mysql"

​
Combined filters

memory search "API design" --project --limit 3

​
How Search Works

​
Hybrid Approach

EchoVault uses a hybrid search strategy:
  1. FTS5 (Full-Text Search): Always active, searches keywords in title, what, why, impact, and tags
  2. Semantic Search: When embeddings are configured, also searches by meaning/similarity
  3. Ranking: Results are scored and ranked by relevance

​
Score Interpretation

  • 0.90-1.00: Excellent match
  • 0.70-0.89: Good match
  • 0.50-0.69: Moderate match
  • Below 0.50: Weak match

​
Output Format

Each result shows:
  • Rank: Position in results [1], [2], etc.
  • Title: Memory title
  • Score: Relevance score (0.00-1.00)
  • Metadata: Category, date (YYYY-MM-DD), project, and source (if set)
  • What: The main description
  • Why: Reasoning (if present)
  • Impact: Consequences (if present)
  • Details indicator: Link to use memory details <id> for extended info

​
When No Results Found

memory search "nonexistent topic"
Output: 
No results found.
Use natural language queries like “why did we…” or “how do we handle…” when semantic search is enabled. The vector embeddings can understand meaning beyond exact keywords.
Semantic search requires an embedding provider configured in config.yaml. Without it, only keyword-based FTS5 search is used (which still works well).

Build docs developers (and LLMs) love

Get started for freeTalk to us