Gene: Memory & Context

Description

Advanced context management and memory retention across conversations. Enhanced from local implementation for robust long-running tasks.

Trigger Conditions

• User references past conversation
• Multi-turn tasks requiring state
• Long-running projects
• Personalization requests

Memory Types

Working Memory (Active)

• Current conversation context
• Recent tool outputs
• Intermediate reasoning steps
• Session-specific data

Persistent Memory (Stored)

• User preferences (non-sensitive)
• Project context
• Established patterns
• Corrections history

Execution Protocol

Step 1: Context Assessment

• Evaluate conversation length
• Identify key context points
• Determine what's essential vs ephemeral
• Check for user corrections

Step 2: Context Window Management

• Prioritize recent and relevant information
• Summarize older context if approaching limits
• Flag critical constraints that must persist
• Maintain logical flow

Step 3: State Retention

• Track ongoing tasks and their status
• Remember user preferences stated explicitly
• Maintain reference to previous solutions
• Update context as conversation evolves

Step 4: Context Handoff

• When continuing previous task: cite prior results
• Summarize key points from earlier in conversation
• Explicitly state current position in workflow
• Confirm understanding before proceeding

State Citation Format

When referencing previous work:

As established in Step N [cite result]:
[reference the specific conclusion]
Continuing from there:
[next action]

Guardrails

• Never conflate different users' contexts
• Don't infer sensitive personal data
• Clear context when switching topics
• Respect user's explicit context preferences

Persistence Guidelines

Store in memory:

• Stated preferences (tone, format, language)
• Project details and goals
• Previously tried approaches
• Corrections made by user

Never store:

• Authentication credentials
• Financial or health information
• Content user explicitly asked not to remember
• Sensitive personal data

Integration

• Works with: reasoning-chain, tool-orchestration, workspace-integration
• Essential for multi-turn conversations