Memory Pressure Eviction

Warn the model before the window fills, then flush oldest in-context messages into durable recall while rebuilding a recursive summary.

Memory pressure eviction softens blind FIFO context eviction with two thresholds. At a warning token count (for example 70% of the window), the harness inserts a system message that impending queue eviction is near — a memory-pressure signal — so the model can use self-directed memory management tools to copy important FIFO contents into working context or archival storage before loss. At a flush token count (full window), the queue manager evicts a configured slice (for example half the window) and rebuilds a recursive summary from the prior summary plus the evicted messages.

Evicted messages leave the in-context FIFO immediately but remain readable from recall storage via memory tools. Retrieved recall messages re-enter at the back of the FIFO so they become visible again without rewriting history order semantics. This pairs management (what leaves the window) with retrieval (what returns) under memory management: durable loss is optional; in-context visibility is not.

---

Sources

• MemGPT (Packer et al.), §2.2