Ralph Memories

Persistent learning system for accumulated wisdom across sessions. Storage: .agent/memories.md.

Memory Injection Configuration

Configure how memories are automatically injected into Ralph's context in ralph.yml:

memories:
  enabled: true                         # Enable/disable memory system
  inject: auto                          # How memories enter context
  budget: 2000                          # Max tokens to inject (0 = unlimited)
  filter:
    types: []                           # Filter by type: pattern, decision, fix, context
    tags: []                            # Filter by tag names
    recent: 0                           # Only memories from last N days (0 = no limit)

Injection Modes

| Mode | Behavior | Use Case | |------|----------|----------| | auto | Ralph prepends relevant memories at start of each iteration | Most workflows - automatic context | | manual | Agent must explicitly run ralph memory search | When you want selective memory access | | none | Memories disabled entirely | Testing or memory-free workflows |

Token Budget

| Setting | Effect | |---------|--------| | 0 | Unlimited memories injected (risky for large memory bases) | | 1500-3000 | Balanced context (recommended) | | 500-1000 | Minimal memory injection for token-constrained workflows |

Filter Configuration

# Only inject pattern memories about API and auth
memories:
  inject: auto
  budget: 1500
  filter:
    types: [pattern]
    tags: [api, auth]
    recent: 30  # Only last 30 days

Prime command equivalents:

# Same filtering via CLI
ralph tools memory prime -t pattern --tags api,auth --recent 30 --budget 1500

# Multiple types (comma-separated)
ralph tools memory prime -t pattern,decision

# Show all memories (no budget limit)
ralph tools memory prime --budget 0

When to Search Memories

Search BEFORE starting work when:

Entering unfamiliar code area → ralph tools memory search "area-name"
Encountering an error → ralph tools memory search -t fix "error message"
Making architectural decisions → ralph tools memory search -t decision "topic"
Something feels familiar → there might be a memory about it

Search strategies:

Start broad, narrow with filters: search "api" → search -t pattern --tags api
Check fixes first for errors: search -t fix "ECONNREFUSED"
Review decisions before changing architecture: search -t decision
Use --all flag to show unlimited results

Working directory option: All ralph tools memory commands support --root <ROOT> to specify working directory (default: current directory).

When to Create Memories

Create a memory when:

You discover how this codebase does things (pattern)
You make or learn why an architectural choice was made (decision)
You solve a problem that might recur (fix)
You learn project-specific knowledge others need (context)

Do NOT create memories for:

Session-specific state (use tasks instead)
Obvious/universal practices
Temporary workarounds

Memory Types

| Type | Flag | Use For | |------|------|---------| | pattern | -t pattern | "Uses barrel exports", "API routes use kebab-case" | | decision | -t decision | "Chose Postgres over SQLite for concurrent writes" | | fix | -t fix | "ECONNREFUSED on :5432 means run docker-compose up" | | context | -t context | "ralph-core is shared lib, ralph-cli is binary" |

Discover Available Tags

Before searching or adding, check what tags already exist:

# See all memories with their tags
ralph tools memory list

# Extract unique tags (grep the file directly)
grep -o 'tags: [^|]*' .agent/memories.md | sort -u

Reuse existing tags for consistency. Common tag patterns:

Component names: api, auth, database, cli
Concerns: testing, performance, error-handling
Tools: docker, postgres, redis

Quick Reference

# Add memory (creates file if needed)
ralph tools memory add "content" -t pattern --tags tag1,tag2

# Add with options
ralph tools memory add "content" -t pattern --tags tag1,tag2 --format quiet

# Search (start broad, narrow with filters)
ralph tools memory search "query"
ralph tools memory search -t fix "error message"
ralph tools memory search --tags api,auth
ralph tools memory search --tags api,auth --all  # Show all results

# List and show
ralph tools memory list
ralph tools memory list -t fix --last 10
ralph tools memory list --format json
ralph tools memory show mem-1737372000-a1b2

# Delete
ralph tools memory delete mem-1737372000-a1b2

# Prime for context injection
ralph tools memory prime --budget 2000
ralph tools memory prime --tags api,auth    # Prime specific tags only
ralph tools memory prime --recent 7         # Only last 7 days
ralph tools memory prime -t pattern         # Prime by type

# Initialize
ralph tools memory init --force             # Overwrite existing file

Output formats: --format {table,json,markdown,quiet}

table: Human-readable table format (default)
json: JSON format for programmatic access
markdown: Markdown format (for prime command)
quiet: ID-only output for scripting

Best Practices

Be specific: "Uses barrel exports in each module" not "Has good patterns"
Include why: "Chose X because Y" not just "Uses X"
One concept per memory: Split complex learnings
Tag consistently: Reuse existing tags when possible

Examples

# Pattern: discovered codebase convention
ralph tools memory add "All API handlers return Result<Json<T>, AppError>" -t pattern --tags api,error-handling

# Decision: learned why something was chosen
ralph tools memory add "Chose JSONL over SQLite: simpler, git-friendly, append-only" -t decision --tags storage,architecture

# Fix: solved a recurring problem
ralph tools memory add "cargo test hangs: kill orphan postgres from previous run" -t fix --tags testing,postgres

# Context: project-specific knowledge
ralph tools memory add "The /legacy folder is deprecated, use /v2 endpoints" -t context --tags api,migration