rig-memory

Conversation memory policies for the Rig agent framework.

rig-core ships the ConversationMemory trait and an in-process InMemoryConversationMemory backend. This crate provides reusable named policies for shaping loaded history before it is sent to the model:

• [NoopMemoryPolicy] — identity policy, useful as a default.
• [SlidingWindowMemory] — keep at most the most recent N messages.
• [TokenWindowMemory] — keep the most recent messages that fit within a token budget supplied by a [TokenCounter].

Both window policies remove the leading prefix through any tool-result messages whose assistant calls were truncated. Results are detected anywhere in a user message, including after text and across intervening system messages. Cleanup stops at the first retained assistant; later paired exchanges remain intact. Whole removed messages are included in apply_with_demoted's demoted prefix in original order, including accompanying text, so demotion hooks lose no content.

Usage

use rig_memory::{InMemoryConversationMemory, SlidingWindowMemory, IntoFilter};

let memory = InMemoryConversationMemory::new()
    .with_filter(SlidingWindowMemory::last_messages(20).into_filter());

For backends other than InMemoryConversationMemory, apply a policy directly:

use rig_memory::{MemoryPolicy, SlidingWindowMemory};

let policy = SlidingWindowMemory::last_messages(20);
let trimmed = policy.apply(loaded_messages)?;

To wrap any backend with a policy and propagate policy errors to the caller (rather than silently degrading to identity on failure), use PolicyMemory:

use rig_memory::{InMemoryConversationMemory, PolicyMemory, SlidingWindowMemory};

let memory = PolicyMemory::new(
    InMemoryConversationMemory::new(),
    SlidingWindowMemory::last_messages(20),
);