The Complete Guide to Paperclip Persistent Memory
Paperclip’s persistent memory is what separates a stateless chatbot from a truly intelligent agent. It lets your agent remember facts, learn preferences, and recall past conversations. Here’s everything you need to know.
Two Types of Memory
Paperclip supports two complementary memory systems that work together.
Key-Value Memory
Think of this as your agent’s notebook — structured facts stored as key-value pairs.
What it stores:
- User preferences (“language: Spanish”, “timezone: UTC-5”)
- Configuration facts (“company_name: Acme Corp”, “plan: enterprise”)
- Session state (“last_order: #12345”, “open_tickets: 3”)
When to use it:
- Exact lookups (“what is X?”)
- Configuration values
- Counters and state tracking
- Structured data that rarely changes
How it works: The agent reads and writes key-value pairs through the memory tool. Values persist across conversations and agent restarts.
Vector Memory
Think of this as your agent’s semantic brain — unstructured knowledge stored as vector embeddings.
What it stores:
- Conversation summaries
- Document passages
- Past decisions and reasoning
- Context from previous interactions
When to use it:
- Fuzzy matching (“what did we discuss about pricing?”)
- Knowledge base queries
- Finding similar past conversations
- Unstructured information retrieval
How it works: Text is converted to vector embeddings and stored in a vector database. When the agent needs to recall something, it searches by semantic similarity, not exact match.
How They Work Together
Most production agents use both:
- Key-value memory stores structured facts the agent needs exactly (“user_id: 123”, “subscription: pro”)
- Vector memory stores conversational context the agent needs approximately (“discussed pricing concerns last week” → retrieves relevant context)
The agent decides which memory type to use based on the query type. You configure the available memory types in your agent settings.
Configuring Memory
Memory Limits
| Plan | KV Memory | Vector Memory |
|---|---|---|
| Starter | 100 keys | 10,000 vectors |
| Pro | Unlimited | 100,000 vectors |
| Scale | Unlimited | Unlimited |
Memory Settings
In your HostAgentes dashboard, configure:
- Auto-summarize — automatically summarize conversations and store in vector memory
- Key extraction — automatically extract facts from conversations into key-value pairs
- Memory scope — per-user, per-session, or global
- Retention — how long to keep memories (default: forever)
Best Practices
1. Don’t Over-Memorize
Not everything needs to be stored. Let the agent decide what’s worth remembering. Configure auto-summarization to capture the gist without storing every word.
2. Use the Right Memory Type
Structured data → key-value. Unstructured context → vector. Mixing them up leads to poor recall.
3. Periodically Clean Up
Old memories can confuse the agent. Set retention policies for stale data. Review key-value pairs monthly and remove outdated facts.
4. Test Memory Recall
Regularly test that your agent can recall important facts. If recall accuracy drops, check your vector memory configuration — embedding models and chunk sizes matter.
5. Scope Appropriately
Per-user memory is usually the right default. Global memory is useful for shared knowledge bases but can create cross-user contamination if not scoped carefully.
Memory on HostAgentes
On HostAgentes, persistent memory is built into the hosting layer. No need to set up Pinecone, Weaviate, or Redis separately. It’s configured through the dashboard and scales automatically with your agent.
Related Posts
Paperclip Governance: Compliance, Policies, and Guardrails
How to implement governance for Paperclip agents — content policies, output filtering, audit trails, compliance frameworks, and responsible AI deployment.
Team Collaboration with Paperclip: Shared Agents and Permissions
How teams collaborate on Paperclip agents — shared dashboards, role-based access, agent ownership, audit logs, and multi-team management.
Cost Optimization Strategies for Paperclip Agents
Reduce your Paperclip agent costs without sacrificing quality. Model selection, token management, caching strategies, and right-sizing your hosting plan.