Memory & Cache¶

Agora has a persistent memory system that lets the model remember information across conversations. Combined with automatic embedding-based caching, it provides a knowledge base that grows with your usage.

Memory Types¶

Active Memory¶

A single, always-on memory context that is included with every API call to the model. Think of it as a sticky note the model always sees.

Use Active Memory for: - Your name, preferences, and background - Project context the model should always know - Standing instructions that apply to all conversations - Facts you're tired of repeating

Example Active Memory content:

User: Newo Ether
Preferences: Prefers Chinese for casual chat, English for technical topics.
Project: Building Agora — a BYOK Android LLM client.
Coding style: Kotlin, Jetpack Compose, MVVM architecture.

Editing Active Memory¶

Go to Settings → Memory
Scroll to Active Memory
Tap Edit Active Memory
Enter your content
Tap Save

The model can also update active memory via tool calls if Access Active Memory is enabled.

Saved Memories¶

A collection of named memory files the model can search, read, create, edit, and delete. Unlike Active Memory (always sent), saved memories are retrieved on demand.

Use Saved Memories for: - Reference material (API docs, config details, commands) - Project-specific notes - Learnings and insights from past conversations - Anything you want the model to recall when relevant

Creating Memories Manually¶

Go to Settings → Memory
Tap Add Memory
Enter:
- Title — descriptive name
- Description — brief summary (used for search matching)
- Content — the full memory content
Tap Create

Model-Created Memories¶

When Access Saved Memories is enabled, the model can create, read, update, and delete memory files via tool calls. This lets the model:

Remember facts you tell it
Save useful code snippets or configurations
Build a knowledge base over time
Clean up outdated information

Memory Permissions¶

Control what the model can access:

Setting	Location	When to Enable
Access Saved Memories	Settings → Memory	You want the model to read/write memory files
Access Active Memory	Settings → Memory	You want the model to update the persistent context
Access Past Conversations	Settings → Conversation Search	You want the model to search chat history

All three default to off. Enable only what you need.

Auto-Cache¶

Auto-caching automatically generates embeddings for new messages as they arrive. This keeps your conversation search index up to date without manual intervention.

Enable Auto-Cache¶

Go to Settings → Conversation Search
Choose an embedding model (if you haven't already — see Embedding / RAG)
Under Caching, toggle Auto-cache new messages

When enabled, every new message (user and model) is automatically embedded and indexed for semantic search.

Manual Caching¶

If auto-cache is off, you can manually cache messages:

Go to Settings → Conversation Search
Tap Cache — computes embeddings for all uncached messages
Progress is shown as a circular indicator

Tap Re-cache to rebuild the entire index from scratch. This deletes all cached embeddings and re-processes every message. Use when: - You've changed embedding models - The cache seems corrupted or outdated - Search results are unexpectedly poor

Warning

Re-caching is irreversible and may take a while depending on your message count and embedding model speed.

Cache Status¶

The embedding model settings show how many messages are cached vs. uncached: - "All N messages cached" — up to date - "X of Y messages not cached" — backlog to process

Memory Tool Calls in Chat¶

When the model uses memory tools, you'll see inline cards:

Tool	Card Text
Look up	"Looked through N saved memories"
Read	"Read [memory name]"
Save	"Saved [memory name]"
Edit	"Updated [memory name]"
Delete	"Removed [memory name]"
Update Active	"Updated active memory"

Tap any card to see the full content that was read or written.

Best Practices¶

Keep Active Memory concise — it's included in every API call, so verbose content wastes tokens
Use descriptive titles for Saved Memories — titles help the model find the right memory
Enable auto-cache if you use conversation search regularly
Re-cache after switching embedding models — different models produce incompatible embeddings