Agentic Context Management

Canonical version: Agentic Context Management.

Elvis Saravia (of DAIR.AI) surfaced a paper that puts a name on something worth keeping in mind: most agent failures in production come from context accumulation, NOT from reasoning defects. The model is fine. Its working memory is drowning.

The paper frames Agentic Context Management around five primitives:

1. Architecting — designing what belongs in context in the first place
2. Ingesting — bringing information in deliberately
3. Scoping — keeping each task's context limited to what that task needs
4. Anticipating — preparing context for what comes next
5. Compacting with consolidation — compressing history without losing what matters

Naive accumulation grows token cost with the square of conversation length (every turn re-reads everything). Proper compaction gets you linear cost, reportedly without fidelity loss; their reference implementation scores 92% on LongMemEval and 93.2% on LoCoMo.

This is the theory underneath Context Engineering, and it converges with the Orchestrator tax argument from a different angle: the scarce resource in agentic systems is a clean Context Window, and it degrades quietly unless you manage it as deliberately as you'd manage memory in a long-running program. It's also the machine-side mirror of Agentic Knowledge Management (AKM): the same discipline of architecting, scoping, and consolidating knowledge, applied to an agent's working memory instead of a human's vault.

References

• Tweet (elvis, @omarsar0): https://x.com/omarsar0/status/2080775791395340313
• Paper: https://arxiv.org/abs/2607.21503

Related

• Agentic Knowledge Management (AKM)
• Context Engineering
• Context Window
• AI Agents
• Orchestrator tax

About Sébastien

Ready to get to the next level?