Research · Sep 6, 2026
Long context is not memory
A million-token context window is still a transcript. Enterprise agents need current state, eviction, and a bind at the checkpoint—not a bigger paste.
Shikhar Mishra · Co-Founder & CEO, EGI
A larger context window can reduce retrieval overhead and preserve more task history. It does not turn the model context into a system of record. For enterprise execution, the window remains a temporary and potentially stale view of operational state.
A context window does not update when the ERP changes, nor does it automatically invalidate a superseded policy. As task history grows, stale tool results can remain available and prior narrative can leak into tool arguments. The execution runtime therefore needs an explicit freshness boundary and a way to select only the state required for the next checkpoint.
Long context vs RAG vs state
RAG selects from a document collection; long context retains more documents or interaction history. Continuously refreshed enterprise state is a different object. EGI maintains normalized pointers and entity handles outside the model, then resolves them at inference and tool-call checkpoints.
- Long-context windows reward dumping. Execution rewards minimum slice size at the gate.
- A context window does not provide record lifecycle semantics; the surrounding system must manage invalidation and eviction.
- Session-wide preload increases the amount of stale or irrelevant state eligible at every subsequent step.
- Model vendors will keep stretching context. That does not make the window a memory architecture.
JIT bind is the enterprise pattern
The hippocampal gate exists because active memory is bounded on purpose. Intent plus tool schema selects a bind target on the organizational index. One record or rule slice enters the window. The call runs. Working memory flushes. That keeps function-specific tool calling independent of session length and lets the same job be reevaluated when the model changes.
For execution workloads, the architectural choice is therefore not simply RAG versus a larger window. The system needs state that can be added, updated, invalidated, and bound at the point of action. Bruce uses the model context as working memory while the authoritative state remains outside it.
Enterprise execution agent
Bruce
The model is a dependency. Bruce is the agent that keeps the job correct—intent through completion on messy, heterogeneous ERPs. The serious work is posting, receiving, and closing.
Related
Enterprise RAG is not an agent
Retrieval-augmented generation answers questions from a corpus. Agents have to complete jobs on live systems. Agentic RAG is still RAG.
The context graph is not a knowledge graph
Agents fail on stale, contradictory, and polluted state—not on a shortage of documents to retrieve.
Why tool calling fails in production
Demos succeed on small schemas and clean worlds. Production is the opposite—and function-specific.