Context Engineering

Prompt engineering asked: what words do I use to get the model to do what I want? Context engineering asks: what does the model need to know, and how do I make sure that knowledge is accurate, current, and structured so the model can reason with it effectively?

The shift sounds subtle. The engineering consequences are not.

#What Changed

Frontier models in 2026 have context windows exceeding one million tokens. The naive response — fill the window with everything relevant — fails on three dimensions simultaneously. Latency: full-context injection produces 10–17 second median time-to-first-token, unusable for real-time applications. Cost: at enterprise scale, injecting full conversation history on every call is financially prohibitive. Accuracy: the “lost in the middle” phenomenon is real and persistent — LLMs systematically fail to recall, weight, and reason over information buried in large context payloads, even with capable models.

The counter-intuitive finding from the LOCOMO benchmark: graph-enhanced selective memory achieves higher accuracy than full-context injection (89.9% vs 72.9%), at lower latency (under 2.6s p95 vs 10–17s), at 90% lower token cost. More context is not better context. Structured, pre-validated, selectively retrieved context is better context.

#The Core Problems

Context stuffing: Treating the context window as a dump for all available information. Token-expensive, latency-degrading, accuracy-reducing. The 2026 anti-pattern that large context windows made worse, not better.

Context rot: Long-running agents accumulate contradictions — superseded policies, changed vendor contracts, updated regulatory requirements. When a context window contains two contradictory policies on the same workflow, the model’s output becomes stochastic. Research in 2026 formalizes this as a survival equation: reasoning accuracy decays exponentially with accumulated contradiction volume. An agent that has been running for months without contradiction resolution is an agent producing increasingly unreliable output.

Identity fragmentation: The same user across mobile, web, and voice channels is treated as three separate entities by standard session-scoped memory. Cross-channel context is siloed. Agents cannot reason about the full relationship without violating session boundaries.

Retrieval blindness: Vector RAG retrieves semantically similar text. It cannot navigate relationships between entities. The connection between an infrastructure decision and its downstream financial outcome requires graph traversal, not embedding proximity.

#What Context Engineering Looks Like in Practice

Six techniques that actually matter in production (per Towards AI’s 2026 analysis):

1. Selective retrieval — graph-based entity traversal over similarity search for relational queries
2. Compression — summarizing historical context before injecting it, reducing token volume without losing semantic density
3. Contradiction detection — flagging conflicting facts before they enter the context window, not after they confuse the model
4. Temporal decay — weighting recent information higher than stale information, with explicit decay curves rather than flat recency windows
5. Identity segmentation — maintaining separate context scopes per principal while resolving cross-scope relationships probabilistically
6. Pre-computation — running context preparation asynchronously so that inference starts from a pre-validated state rather than raw retrieval

#Where Mumega Sits

The Metabolism Layer is context engineering at the infrastructure layer. It runs asynchronously during idle periods — what the OSF research calls “cognitive sleep” — scanning the Mirror graph, detecting contradictions via LLM-based pairwise comparison, resolving them, and applying decay to stale facts. By the time a Mumega agent is invoked for a live task, the context has already been prepared: contradictions resolved, stale facts weighted down, entity relationships updated.

The agent receives a compressed, pre-validated payload, not a raw accumulation. It boots with current situational accuracy in milliseconds. The Metabolism Layer is what makes that possible without requiring the inference call to do the work.

The Amrita Score handles the identity fragmentation problem: probabilistic cross-session entity resolution that calculates a confidence threshold before linking fragmented interaction records into a unified graph.

Both are infrastructure decisions, not prompt decisions. They cannot be replicated by writing better system prompts.

#Related

• AI Agent Memory — The graph memory architecture that context engineering runs on
• Multi-Agent Orchestration — How shared context infrastructure eliminates inter-agent update overhead
• Agentic Governance & Security — Why audit records cannot be purged during contradiction resolution — the memory layer’s constraint