grey-haven-memory-systems

# Memory Systems Skill

Design and implement long-term memory systems for AI agents.

## The Context-Memory Spectrum

Memory exists on a spectrum from ephemeral to permanent:

```
Ephemeral ◄────────────────────────────────────► Permanent

Context Window    Short-term    Long-term    Knowledge
(disappears)      Cache         Memory       Base
                  (session)     (weeks)      (forever)
```

### When to Use What

| Memory Type | Duration | Use Case |
|-------------|----------|----------|
| Context window | Single turn | Immediate task context |
| Short-term cache | Session | Conversation history |
| Long-term memory | Weeks/months | User preferences, learnings |
| Knowledge base | Permanent | Facts, documentation, procedures |

## Memory Architecture Options

### 1. Vector RAG (Retrieval-Augmented Generation)

Store embeddings, retrieve by semantic similarity.

**Pros**:
- Simple to implement
- Works well for document retrieval
- Scales to millions of documents

**Cons**:
- No relationships between items
- Recency bias (older memories fade)
- Can retrieve irrelevant but similar content

**Best for**: Document search, FAQ systems, code search

### 2. Knowledge Graphs

Store entities and relationships explicitly.

**Pros**:
- Captures relationships
- Supports reasoning
- No similarity confusion

**Cons**:
- Complex to build and maintain
- Requires structured data
- More expensive queries

**Best for**: Domain modeling, reasoning tasks, complex queries

### 3. Temporal Knowledge Graphs

Knowledge graphs with time-based relationships.

**Pros**:
- Tracks how knowledge evolves
- Supports "as of" queries
- Captures causality

**Cons**:
- Most complex option
- Storage grows over time
- Query complexity

**Best for**: Historical analysis, change tracking, audit trails

### 4. Hybrid Approaches

Combine vector + graph for best of both:

```
Query ──▶ Vector Search ──▶ Top K candidates
              │
              ▼
          Graph Traversal ──▶ Related entities