Context Management Patterns for Long-Running AI Agents

Long-running AI agent sessions suffer from context collapse — a set of failure modes that emerge as conversation context grows: - **Constraint loss**: earlier instructions forgotten mid-session - **Redundant work**: agent repeats identical tool calls it already executed - **Contradictory decisions**: later outputs conflict with earlier reasoning - **Dilution**: instructions technically present but buried, degrading performance 14–85% (the "lost in the middle" effect) The root cause is not memory capacity but context management. Tool outputs accumulate with no distinction between current and stale. Old constraints sit alongside new work with no signal for relevance. **Episode-based storage** addresses this by storing complete turns (user input + tool calls + results + response) as atomic episodes. Nothing gets summarized or compressed — it gets indexed and moved out of active context but remains fully retrievable. A sliding window keeps recent episodes active while older ones are indexed for retrieval. This preserves the full reasoning chain rather than just the conclusion. **AGNES (agglomerative nesting)** automatically groups related episodes into a hierarchical tree structure without predefined categories. Topics emerge organically from the content (e.g., travel → Japan → visa docs). This scales without manual curation and makes retrieval decisions explainable through full path tracing. **Multi-branch beam search** queries across multiple branches of the hierarchy simultaneously, scoring by both relevance and recency. This prevents the lost-in-the-middle problem where important older context gets buried. A query like "debugging deployment" can hit both devops→docker AND elixir→phoenix→deployment, catching cross-domain insights that single-path search misses. The core insight from Contexto (by Ekai Labs): **curation beats compression**. Don't summarize and lose information — index it, cluster it, and retrieve selectively. Keep raw data intact and do consolidation as a view layer on top, not a destructive transform. Source: Contexto by Ekai Labs, Apache 2.0 licensed, TypeScript/OpenClaw plugin.