slm-recall

name	slm-recall
description	Search SuperLocalMemory for relevant facts, decisions, and past context. Use when the user asks to recall, search, find, or retrieve stored information. Invokes 5-channel retrieval with LightGBM reranking via MCP.
version	3.4.23
license	AGPL-3.0-or-later
compatibility	SuperLocalMemory v3.4.22 — MCP (preferred) or CLI fallback
attribution	{"creator":"Varun Pratap Bhardwaj","role":"Solution Architect & Creator","project":"SuperLocalMemory v3"}

SuperLocalMemory: Recall

Search and retrieve memories using semantic similarity, knowledge graph relationships, and full-text search.

Usage

slm recall "<query>" [--limit N] [--min-score 0.0-1.0] [--tags tag1,tag2] [--project name]

Examples

Example 1: Basic Search

slm recall "FastAPI"

Output:

🔍 Search Results (3 found)

[ID: 42] Score: 0.85
We use FastAPI for REST APIs
Tags: python, backend, api
Project: myapp
Created: 2026-02-05 14:23

[ID: 38] Score: 0.72
FastAPI is faster than Flask for high-throughput APIs
Tags: performance, python
Project: default
Created: 2026-02-01 09:15

[ID: 29] Score: 0.68
Async endpoints in FastAPI improve concurrency
Tags: async, fastapi, python
Project: myapp
Created: 2026-01-28 11:42

Example 2: Limited Results

slm recall "authentication" --limit 3

Returns: Top 3 most relevant results

Example 3: Minimum Relevance Score

slm recall "React hooks" --min-score 0.7

Only returns results with relevance score ≥ 0.7

Example 4: Filter by Tags

slm recall "database" --tags postgresql,performance

Only searches memories tagged with specified tags

Example 5: Filter by Project

slm recall "API design" --project myapp

Only searches memories in specified project

Arguments

Argument	Type	Required	Default	Description
<query>	string	Yes	-	Search query
--limit	integer	No	10	Max results to return
--min-score	float	No	0.3	Minimum relevance (0.0-1-tags`
--project	string	No	None	Filter by project

Search Methods Used

SuperLocalMemory uses 3 search methods simultaneously and merges results:

1. Semantic Search (TF-IDF)

• Converts query to vector
• Finds similar content vectors
• Best for: Conceptual matches

Example:

Query: "authentication patterns"
Matches: "JWT tokens", "OAuth flow", "session management"

2. Knowledge Graph Traversal

• Finds related memories via graph edges
• Discovers connected concepts
• Best for: Related information

Example:

Query: "FastAPI"
Graph finds: "REST API" → "JWT auth" → "token refresh"

3. Full-Text Search (FTS5)

• Exact keyword matching
• Fast for known terms
• Best for: Specific phrases

Example:

Query: "PostgreSQL 15"
Finds: Exact mentions of "PostgreSQL 15"

Relevance Scores

Score range: 0.0 - 1.0

Score	Meaning
0.9 - 1.0	Excellent match (almost exact)
0.7 - 0.9	Strong match (very relevant)
0.5 - 0.7	Good match (related)
0.3 - 0.5	Weak match (loosely related)
< 0.3	Poor match (filtered out by default)

Factors affecting score:

• Keyword overlap
• Semantic similarity
• Graph distance
• Recency (newer = slight boost)
• Importance level
• Your usage patterns (results improve automatically over time)

Advanced Usage

Natural Language (in AI chat)

Most AI assistants automatically invoke this skill when you ask:

• "What did we decide about..."
• "Recall information about..."
• "Search for..."
• "What do we know about..."

Example in Cursor/Claude:

You: "What did we decide about authentication?"
AI: [Automatically invokes slm-recall skill]
Found 3 memories about JWT tokens and OAuth...

Combined with Other Skills

1. Recall then remember:

# Find existing memories
slm recall "API design"

# Add new related memory
slm remember "New API versioning strategy: use /v2/ prefix" --tags api,versioning

2. Recall then build graph:

# Find memories
slm recall "performance"

# Rebuild graph to discover new connections
slm build-graph

Scripting & Automation

Find and export:

# Search and save to file
slm recall "security" --min-score 0.7 > security-notes.txt

# Count memories matching query
slm recall "python" --limit 999 | grep "^\\[ID:" | wc -l

Regular reminders:

# Daily standup helper (cron job)
#!/bin/bash
echo "Yesterday's decisions:"
slm recall "decided" --limit 5

echo -e "\nCurrent blockers:"
slm recall "blocked" --tags critical --limit 3

Troubleshooting

"No memories found"

Causes:

1. No memories matching query
2. Min-score too high
3. Wrong project filter

Solutions:

# Lower minimum score
slm recall "query" --min-score 0.1

# Remove filters
slm recall "query"  # No project/tag filters

# Check what memories exist
slm list --limit 20

"Search too slow"

Causes:

• Large database (10,000+ memories)
• Complex query
• Knowledge graph not optimized

Solutions:

# Rebuild indexes
slm build-graph

# Use filters to narrow search
slm recall "query" --project myapp --tags specific-tag

# Increase min-score (fewer results = faster)
slm recall "query" --min-score 0.7

"Results not relevant"

Causes:

• Query too vague
• Need to add more context

Solutions:

# Be more specific
❌ slm recall "it"
✅ slm recall "authentication system"

# Use multiple keywords
✅ slm recall "FastAPI JWT authentication"

# Use tags to filter
✅ slm recall "performance" --tags database

Output Formats

Standard Format (Default)

🔍 Search Results (3 found)

[ID: 42] Score: 0.85
Content preview...
Tags: tag1, tag2
Project: myapp
Created: 2026-02-05

Programmatic Use

# JSON output (for scripts)
slm recall "query" --format json
# {"results": [{"id": 42, "content": "...", "score": 0.85}, ...]}

# CSV output
slm recall "query" --format csv
# id,content,score,tags,project,created_at
# 42,"Content...",0.85,"tag1,tag2",myapp,2026-02-05

Performance Benchmarks

Database Size	Search Time	Notes
100 memories	~100ms	Instant
1,000 memories	~500ms	Fast
10,000 memories	~1.5s	Acceptable
50,000 memories	~5s	Consider filtering

Optimization tips:

• Use --min-score to filter early
• Use --tags or --project to narrow search
• Rebuild graph periodically: slm build-graph

Notes

• Multi-method: Combines semantic, graph, and keyword search
• Ranked results: Best matches first
• Cross-tool: Same results in Cursor, ChatGPT, Claude, etc.
• Privacy: All search happens locally
• Real-time: Database updates reflected immediately

Related Commands

• slm remember "<content>" - Save a new memory
• slm list - List recent memories (no search)
• slm status - Check memory count and graph stats
• slm build-graph - Optimize search performance

---

Created by: Varun Pratap Bhardwaj (Solution Architect) Project: SuperLocalMemory V2 License: MIT (see LICENSE) Repository: https://github.com/varun369/SuperLocalMemoryV2

Open source doesn't mean removing credit. Attribution must be preserved per MIT License terms.

V3 Update (v3.4.22)

Preferred: Use the recall MCP tool directly — it uses 5-channel retrieval (semantic, BM25, entity-graph, temporal, Hopfield) with LightGBM reranking and feeds the learning loop.

recall(query="your query", limit=10, session_id="<current_session>")

CLI fallback (also valid, no MCP required):

slm recall "<query>" [--limit N]

Always pass session_id when using MCP so engagement signals feed the ranker.