-
Notifications
You must be signed in to change notification settings - Fork 647
Description
Priority: Low
Depends on: #1679
Problem Statement
Content removed from the active context window is stored in SessionManager but never retrieved. Once content leaves the active window, it cannot come back even if it becomes relevant later. The agent passively receives whatever fits in the window rather than actively managing its own memory.
Proposed Solution
Give the agent meta-tools to navigate its own context, following the MemGPT pattern where the agent manages its own memory.
@tool
def populate_context(...) -> None:
"""Find and load relevant context into the active window."""
...The agent can request context, and the framework handles retrieval from session storage and insertion into the active window. The agent can also strategically order retrieved context, placing high-relevance content at attention-favored positions (beginning and end of context per "Lost in the Middle" research).
Use Case
- Long-running conversations where early context becomes relevant again
- Agents that need to recall specific details from earlier in the conversation
- Strategic context ordering for better model attention
Additional Context
Part of the Context Management epic, Track 1: Conversation Context. Depends on #1679 (Bridge ConversationManager and SessionManager).