Cross Project Learning

Cross-Project Learning

Loki Mode's memory system that learns from every session and applies insights to future projects.

---

Overview

Cross-project learning captures three types of knowledge:

Type	Description	Example
Patterns	Reusable approaches	"Use JWT with refresh tokens for auth"
Mistakes	Errors to avoid	"Don't store secrets in localStorage"
Successes	What worked well	"TDD reduced bugs by 60%"

---

How It Works

Learning Extraction

At the end of each session, Loki Mode:

1. Analyzes CONTINUITY.md for patterns
2. Extracts learnings from session logs
3. Deduplicates using MD5 hashing (reduces 71% duplicates)
4. Stores in JSONL format for efficient append

Learning Application

At the start of each session, Loki Mode:

1. Loads relevant learnings from global memory
2. Applies patterns to current context
3. Avoids known mistakes
4. Reuses successful approaches

---

Storage Location

~/.loki/learnings/
  patterns.jsonl     # Reusable patterns
  mistakes.jsonl     # Errors to avoid
  successes.jsonl    # Successful approaches

Storage Format

Each file uses JSONL (JSON Lines) format:

{"version":"1.0","created":"2026-02-02T12:00:00Z"}
{"description":"Always validate user input at API boundaries","project":"auth-service","category":"security","timestamp":"2026-02-02T12:30:00Z"}
{"description":"Use connection pooling for database connections","project":"data-api","category":"performance","timestamp":"2026-02-02T14:00:00Z"}

---

CLI Commands

List Learnings

# List all learnings
loki memory list

# List specific type
loki memory show patterns
loki memory show mistakes
loki memory show successes

# With limit
loki memory show patterns --limit 10

Search Learnings

# Search across all types
loki memory search "authentication"

# Output as JSON
loki memory search "database" --format json

Statistics

# View statistics
loki memory stats

Output:

Cross-Project Learnings Statistics

By Category:
  patterns:  25
  mistakes:  10
  successes: 15
  Total:     50

By Project:
  auth-service:  20
  data-api:      15
  frontend:      10
  unknown:        5

Export/Import

# Export all learnings
loki memory export ./learnings-backup.json

# Import learnings
loki memory import ./learnings-backup.json

Clear Learnings

# Clear specific type
loki memory clear patterns
loki memory clear mistakes

# Clear all
loki memory clear all

Deduplicate

# Remove duplicate entries
loki memory dedupe

---

API Endpoints

Get Summary

curl http://localhost:57374/memory

Response:

{
  "patterns": 25,
  "mistakes": 10,
  "successes": 15,
  "location": "/Users/you/.loki/learnings"
}

Get Learnings by Type

curl "http://localhost:57374/memory/patterns?limit=10"

Response:

{
  "type": "patterns",
  "entries": [
    {
      "description": "Use JWT with refresh tokens",
      "project": "auth-service",
      "timestamp": "2026-02-02T12:00:00Z"
    }
  ],
  "total": 25,
  "limit": 10,
  "offset": 0
}

Search

curl "http://localhost:57374/memory/search?q=authentication"

Clear

curl -X DELETE http://localhost:57374/memory/patterns

---

Configuration

Environment Variables

Variable	Default	Description
LOKI_MEMORY_DIR	~/.loki/learnings	Storage location
LOKI_MEMORY_ENABLED	true	Enable/disable learning
LOKI_MEMORY_DEDUPE	true	Auto-deduplicate

Config File

# ~/.config/loki-mode/config.yaml
memory:
  enabled: true
  directory: ~/.loki/learnings
  dedupe: true
  max_entries_per_type: 1000

---

CONTINUITY.md Format

Loki Mode extracts learnings from CONTINUITY.md:

# Session Continuity

## Patterns Discovered
- Always validate user input at API boundaries
- Use connection pooling for database connections

## Mistakes Made
- Forgot to handle null case in user lookup
- Didn't set up proper error boundaries

## What Worked Well
- TDD approach caught 3 bugs early
- Component-first design simplified testing

---

Best Practices

Writing Good Learnings

1. Be specific - "Use bcrypt with cost factor 12" vs "Hash passwords"
2. Include context - "For REST APIs, use..." vs "Use..."
3. Keep it actionable - Focus on what to do/avoid

Managing Learnings

1. Review periodically - Remove outdated learnings
2. Deduplicate regularly - Run loki memory dedupe
3. Export backups - Keep backups of valuable learnings
4. Share selectively - Review before sharing with team

---

Knowledge Compounding (v5.30.0)

In addition to raw JSONL learnings, Loki Mode can extract structured solutions -- curated, categorized knowledge with YAML frontmatter that feeds back into future planning.

How It Differs from Raw Learnings

Aspect	Raw Learnings (JSONL)	Structured Solutions (MD)
Format	JSON lines, append-only	Markdown with YAML frontmatter
Structure	Flat description field	title, tags, symptoms, root_cause, prevention
Storage	~/.loki/learnings/*.jsonl	~/.loki/solutions/{category}/*.md
Retrieval	Keyword grep	Tag + symptom matching, relevance scoring
Categories	patterns/mistakes/successes	security/performance/architecture/testing/debugging/deployment/general

Solution File Format

---
title: "Connection pool exhaustion under load"
category: performance
tags: [database, pool, timeout, postgres]
symptoms:
  - "ECONNREFUSED on database queries under load"
root_cause: "Default pool size of 10 insufficient"
prevention: "Set pool size to 2x concurrent connections"
confidence: 0.85
source_project: "auth-service"
created: "2026-02-09T12:00:00Z"
applied_count: 0
---

## Solution
[Detailed explanation of fix]

## Context
[How this was discovered]

When Solutions Are Created

• After VERIFY passes with novel insight (automatic via COMPOUND phase)
• When fixing a non-obvious bug (root cause analysis)
• When discovering a reusable pattern
• When hitting a pitfall worth documenting
• Manually via loki compound run

CLI Commands

# List solutions by category
loki compound list

# Show solutions in a category
loki compound show security

# Search across all solutions
loki compound search "authentication"

# Manually trigger compounding from session learnings
loki compound run

# View statistics
loki compound stats

Solution Loading

At the start of each session, Loki Mode:

1. Scans ~/.loki/solutions/ subdirectories
2. Reads YAML frontmatter from each solution file
3. Matches tags and symptoms against current task context
4. Injects top 3 relevant solutions into the REASON phase context

---

See Also

• API Reference - Memory API endpoints
• CLI Reference - Memory CLI commands
• Architecture - Memory system architecture