Overview
/
Decision Tree
Last updated

Should You Use Papr?

Quick decision guide to help you choose the right memory approach.

For the full detailed comparison and implementation code, continue to Why Papr.

Decision Flow

Start here ↓

Do you need memory across sessions?
├─ No → Session state (no DB needed)
└─ Yes ↓

Will users query with different words than stored?
├─ No → SQLite + FTS5 (keyword only)
└─ Yes ↓

Do you need relationships between data?
├─ No → Vector DB (semantic only)
└─ Yes ↓

Do you want graph-aware retrieval only, or memory that gets better with usage?
├─ Graph-only is enough → DIY vector + graph stack
└─ Yes → Use Papr ✓

What Each Choice Means

Session State Only

When: Single conversation, no persistence needed
Example: Chatbot that doesn't remember between sessions
Next: No memory layer needed

SQLite + FTS5

When: Multi-session, exact keyword matching sufficient
Limitation: No semantic search (misses "refund" when you stored "return")
Next: SQLite FTS5 docs

Vector DB

When: Need semantic search, no relationships
Limitation: Can't connect entities (Person → Project → Bug)
Next: Choose Pinecone, Weaviate, or similar

DIY Hybrid Stack

When: You only need vector + graph retrieval and can operate your own stack
Reality: 6+ systems to manage, 2-3 months build, 0.5-1 FTE ongoing
Next: Why Papr

Papr

When: Need semantic + relationships + consolidation
Reality: 15 min to prototype, 0 FTE maintenance
Key difference: Memory improves from usage (predictive), not frozen
Next: Quick Start

Why Relationships Matter

Vector-only example:

Query: "What's the status of the auth bug?"
Returns: Text fragments with "auth" and "bug"
Problem: Which bug? Who's working on it? Related PRs?

With graph:

Query: "What's the status of the auth bug?"
Returns:
  - Bug ticket with status
  - Assignee and priority
  - Related PRs and code
  - Discussions and decisions
All connected, not just similar text

Why Predictive Memory Matters (Beyond Vector + Graph)

Vector + graph gives you:

  • Better semantic matching
  • Relationship traversal
  • Richer retrieval than keyword-only

Predictive memory adds:

  • Behavioral learning from usage patterns over time
  • Anticipatory context preloading for lower latency (when cached)
  • Ranking that improves as retrieval logs and interaction history grow

In short: vector + graph improves what you can retrieve now; predictive memory improves how retrieval quality evolves over time.

Which Papr API to Use

Once you've chosen Papr, pick your starting point:

By Input Type

  • Chat/conversations → /v1/messages (quickstart)
  • Documents (PDFs, Word) → /v1/document (quickstart)
  • Postgres/SQL records → /v1/memory manual mode (quickstart)
  • Agent learnings → /v1/memory auto mode (quickstart)

By Goal

Next Steps

Ready to start:

Need more details:

Previous page
Next page