Memory Recording

Memory Recording Skill

Use this skill to persist valuable knowledge acquired during tasks. This creates "mid-term memory" that bridges the gap between training data and future sessions.

When to Record

Record memories at these trigger points:

1. End of task: After completing any significant work
2. End of session: Before the conversation ends
3. Before context compaction: When context is about to be summarized
4. After important discoveries: When learning something that should persist

What to Record

SHOULD Record

• Architectural decisions and their rationale
• Design patterns adopted for this project
• Bug analyses and root causes discovered
• Configuration details that required investigation
• User preferences and working style observations
• Project conventions (naming, structure, testing approach)
• Integration details (APIs, services, dependencies)
• Troubleshooting findings that took effort to discover

Should NOT Record

• Ephemeral task details (temporary file paths, session-specific variables)
• Information already in version control (commit messages capture this)
• Trivial facts easily re-discovered (standard library usage)
• Sensitive data (credentials, API keys, personal information)

Recording Procedure

Step 1: Identify New Knowledge

Ask yourself:

1. What did I learn that I didn't know at task start?
2. What decisions were made and why?
3. What would be valuable to know in a future session?
4. What took significant effort to discover or understand?

Step 2: Check for Existing Entities

Before creating new entities, search for existing ones to update:

memento_search_nodes(query: "<concept you want to record>")

If a relevant entity exists, use memento_add_observations to append new information rather than creating duplicates.

Step 3: Structure the Entity

Entity Name

Use descriptive, human-readable names:

• Good: "Rizer Database Connection Pooling Strategy"
• Good: "PR #42 Authentication Bug Analysis"
• Bad: "db_config" (too terse)
• Bad: "stuff I learned today" (not descriptive)

Naming patterns:

• <Project> <Topic> <Type> - "Rizer API Error Handling Pattern"
• <Identifier> <Description> - "ADR-0001 Event Sourcing Decision"
• <Topic> <Date Context> - "CI Pipeline Fix January 2026"

Entity Type

Use descriptive snake_case or Title Case types. Common types observed:

Choose the most specific type that fits, or create a new descriptive type.

First Observation (Project Association)

ALWAYS include project metadata as the first observation:

Project: rizer | Path: /home/jwilger/projects/rizer | Scope: PROJECT_SPECIFIC

For general knowledge not tied to a specific project:

Scope: GENERAL | Topic: <general topic area>

Subsequent Observations

Each observation should be a complete, self-contained fact:

• Good: "Uses connection pooling with max 10 connections per database"
• Good: "Decision made on 2026-01-20 after performance testing showed 3x improvement"
• Bad: "pooling" (too terse, lacks context)
• Bad: "See above" (not self-contained)

Include:

• Specific technical details
• Rationale for decisions (the "why")
• Dates when relevant
• File paths for code-related knowledge
• Constraints or limitations discovered

Step 4: Create or Update Entity

Creating New Entity

memento_create_entities(entities: [{
  name: "Rizer Database Connection Strategy",
  entityType: "architecture_decision",
  observations: [
    "Project: rizer | Path: /home/jwilger/projects/rizer | Scope: PROJECT_SPECIFIC",
    "Decision: Use sqlx with connection pooling, max 10 connections",
    "Rationale: Testing showed 3x throughput improvement over single connection",
    "Date: 2026-01-20",
    "Configuration in src/db/pool.rs with PgPoolOptions",
    "Environment variable DATABASE_MAX_CONNECTIONS controls pool size"
  ]
}])

Updating Existing Entity

memento_add_observations(observations: [{
  entityName: "Rizer Database Connection Strategy",
  contents: [
    "Update 2026-01-25: Added connection timeout of 30s after intermittent failures",
    "New config option DATABASE_CONNECT_TIMEOUT added"
  ]
}])

Step 5: Create Relations (When Meaningful)

If the new knowledge connects to existing entities, create relations:

memento_create_relations(relations: [{
  from: "Rizer Database Connection Strategy",
  to: "Rizer Performance Optimization Plan",
  relationType: "implements",
  strength: 0.9,
  confidence: 1.0
}])

Common relation types:

• implements - This implements/satisfies that
• depends_on - This requires that
• relates_to - General association
• supersedes - This replaces that decision
• caused_by - This issue was caused by that

Recording Checklist

Before finishing a task/session, verify:

• Did I make any decisions that should be recorded?
• Did I discover anything non-obvious about the codebase?
• Did I establish any patterns or conventions?
• Did I encounter and resolve any issues worth remembering?
• Did I learn any user preferences or project requirements?

Example Recording Session

After implementing authentication for rizer:

1. Search for existing: memento_search_nodes("rizer authentication")
   → No existing entities found

2. Create new entity:
   memento_create_entities(entities: [{
     name: "Rizer Authentication Architecture",
     entityType: "architecture_decision",
     observations: [
       "Project: rizer | Path: /home/jwilger/projects/rizer | Scope: PROJECT_SPECIFIC",
       "Decision: JWT-based authentication with refresh tokens",
       "Library: jsonwebtoken crate for JWT handling",
       "Token expiry: 15 minutes access, 7 days refresh",
       "Storage: Refresh tokens stored in PostgreSQL with user_sessions table",
       "Rationale: Stateless access tokens reduce DB load, refresh tokens allow revocation",
       "Security: Tokens signed with RS256, keys rotated monthly",
       "Implementation date: 2026-01-20",
       "Key files: src/auth/jwt.rs, src/auth/middleware.rs, src/db/sessions.rs"
     ]
   }])

3. Create relation to existing user entity:
   memento_create_relations(relations: [{
     from: "Rizer Authentication Architecture",
     to: "Rizer User Management",
     relationType: "implements",
     strength: 0.8,
     confidence: 1.0
   }])

Important Notes

• Be specific: Vague memories are useless memories
• Include rationale: The "why" is often more valuable than the "what"
• Date significant decisions: Helps understand evolution over time
• Don't over-record: Focus on knowledge that provides future value
• Update don't duplicate: Search first, add observations to existing entities