[FEATURE] Tiered Memory (MemGPT-inspired)

[dbschmigelski]

Priority: Low
Depends on: #1679, #1682

Problem Statement

Current conversation managers treat context reduction as a one-way operation. Sliding window drops messages permanently. Summarization compresses them into a lossy summary. In both cases, once content leaves the active window, it cannot come back. These approaches work for short conversations but struggle to maintain trajectory alignment over long-running sessions.

Proposed Solution

Implement OS-style tiered memory with intelligent paging between three tiers, inspired by MemGPT:

• Main context (active window) — what the model sees right now
• Recall memory (recent history) — quickly retrievable recent content
• Archival memory (long-term storage) — persistent, searchable storage

Content moves between tiers based on relevance. The existing SessionManager provides the storage layer; this feature adds the intelligent paging and retrieval logic.

The key insight is that Strands already has the storage layer but lacks the retrieval path. A first-class implementation works across any model provider and storage backend, providing provider independence.

Use Case

• Long-running agent sessions that need to maintain trajectory alignment
• Conversations where early context becomes relevant again later
• Production agents that run for hours or days

Additional Context

Part of the Context Management epic, Track 1: Conversation Context. This is the capstone feature. Depends on #1679 (Bridge ConversationManager/SessionManager) and #1682 (Context Navigation Meta-tools).