Memory Manager

Memory Manager

Goal

Maintain persistent context across all sessions through a three-layer memory system.

Memory Layers

Layer 1: Daily Notes

Location: memory/daily/YYYY-MM-DD.md Purpose: Raw session logs and interaction history Update: After each significant interaction

Layer 2: Long-Term Memory

Location: memory/MEMORY.md Purpose: Curated knowledge, preferences, decisions Update: When learning important context

Layer 3: Semantic Index (Optional)

Location: memory/embeddings/ Purpose: Vector search for conceptual retrieval Update: Periodically index memory content

Process

Writing to Daily Notes

After significant interactions, append:

## HH:MM - [Session Type]

**User:** [Request summary]
**Agent:** [Action taken]

### Details
- Key decisions made
- Files modified
- Outcomes achieved

### Follow-up
- Pending items
- Questions to revisit

Updating Long-Term Memory

When learning important context:

## [Category]

### [Topic]
- Key fact or preference
- Supporting details
- Date learned: YYYY-MM-DD

Categories to maintain:

• User Preferences: Communication style, tool preferences, work patterns
• Project Context: Tech stack, architecture decisions, key files
• Important Decisions: What was decided and why
• Learned Patterns: Shortcuts, conventions, recurring needs

Memory Retrieval

Before responding to any request:

1. Check if MEMORY.md has relevant context
2. Scan recent daily notes (today, yesterday)
3. For complex queries, search semantically if available

Memory Consolidation

Periodically (or when daily notes grow large):

1. Review recent daily notes
2. Extract patterns and important facts
3. Update MEMORY.md with distilled knowledge
4. Archive old daily notes if needed

Memory Templates

Initial MEMORY.md

# Agent Memory

*Last updated: YYYY-MM-DD*

## User Profile

### Preferences
- [To be learned]

### Work Patterns
- [To be learned]

## Current Context

### Active Project
- **Name:** [Project name]
- **Stack:** [Technologies]
- **Status:** [Current phase]

### Key Files
- [To be discovered]

## Important Decisions

*Decisions will be logged as they're made*

## Learned Patterns

*Patterns will be captured from interactions*

Daily Note Template

# YYYY-MM-DD

## Summary
[Brief overview of the day's work]

## Sessions

### HH:MM - [Type]
[Session details]

## Key Learnings
- [What was learned]

## Pending Items
- [ ] [Items to follow up]

Auto-Flush Protocol

When context window approaches 80% capacity:

1. Trigger flush before compaction
2. Extract key information from current context:
   • Decisions made
   • Preferences expressed
   • Important facts mentioned
3. Write to MEMORY.md
4. Log summary to daily notes
5. Allow compaction to proceed

Memory Hygiene

Do

• Update memory after learning something important
• Be specific and factual in entries
• Include dates for time-sensitive information
• Cross-reference related entries

Don't

• Store sensitive credentials in memory files
• Log every trivial interaction
• Duplicate information across layers
• Let memory become stale without review

Integration with Other Skills

With skill-creator

When a new skill is created, log to memory:

## Learned Patterns

### Skills Acquired
- **[skill-name]** (YYYY-MM-DD): [what it does]

With heartbeat

During heartbeat checks, consult memory for:

• User's active hours and preferences
• Pending items that need attention
• Context for proactive actions

Commands

• "Remember this" - Explicitly save current context to MEMORY.md
• "What do you know about X" - Retrieve from memory
• "Update memory" - Trigger consolidation
• "Show memory" - Display current MEMORY.md

References

• See AUTONOMOUS_BOOTUP_SPEC.md for architecture
• See skill-creator for skill persistence
• See heartbeat-manager for scheduled memory tasks