Memory

Memory determines how your agent remembers past conversations. TraitClaw provides three built-in implementations and a trait for building custom ones.

Memory Backends

Backend	Persistence	Use Case
InMemoryMemory	❌ Process lifetime	Prototyping, stateless APIs
SqliteMemory	✅ Disk	Production single-node
CompressedMemory	✅ + Compression	Long conversations, cost optimization

InMemoryMemory (Default)

Zero-config, works out of the box:

// This is the default — no setup needed
let agent = Agent::builder()
    .provider(provider)
    .build()?;

// Explicit
let agent = Agent::builder()
    .provider(provider)
    .memory(InMemoryMemory::new())
    .build()?;

SqliteMemory

Persistent conversations with full-text search:

use traitclaw_memory_sqlite::SqliteMemory;

let memory = SqliteMemory::new("./agent.db")?;

let agent = Agent::builder()
    .provider(provider)
    .memory(memory)
    .session("user-123")  // Session isolation
    .build()?;

// Conversations persist across restarts
let output = agent.run("Remember: my name is Alice").await?;
// ... restart app ...
let output = agent.run("What's my name?").await?;
// "Your name is Alice!"

CompressedMemory

For long conversations, CompressedMemory automatically summarizes older messages to save tokens:

use traitclaw_core::CompressedMemory;

let memory = CompressedMemory::new(
    inner_memory,
    compression_provider,  // LLM used for summarization
)
.max_tokens(4000)          // Compress when context exceeds this
.keep_recent(5);           // Always keep the 5 most recent messages

The Memory Trait

Build custom backends for Redis, PostgreSQL, DynamoDB, etc.:

#[async_trait]
pub trait Memory: Send + Sync {
    /// Retrieve all messages for a session
    async fn get_messages(&self, session_id: &str) -> Result<Vec<Message>>;

    /// Add a message to a session
    async fn add_message(&self, session_id: &str, message: Message) -> Result<()>;

    /// Clear all messages for a session
    async fn clear(&self, session_id: &str) -> Result<()>;
}