Context window
The context window is the maximum number of tokens an LLM can hold in a single inference call. It bounds how much conversation history, retrieved memory, and tool output the model can reason over in one turn. Modern frontier models support 100K-200K+ token windows.
Why it matters
Every LLM call is bounded by the context window. The agent's working-conversation history, any memory retrieved from long-term storage, the tool-call output, and the model's response all have to fit. For an agent running multi-turn conversations with tool calls, context budget is a constant operational concern.
Two failure modes when the context window is mismanaged. First, the agent forgets earlier in the conversation because old turns got dropped. Second, the agent stops being able to read tool output because the tool output is too large.
The engineering response is layered: working memory holds the current conversation; long-term memory is retrieved on demand via embedding search; tool outputs that exceed budget are summarized before being passed back to the model. Done well, the user never notices the budget. Done poorly, the agent's coherence collapses on long sessions.
Frontier model context windows have grown from a few thousand tokens (early GPT-3) to 100K-200K+ tokens (Claude, GPT-4 long-context variants) in 2024-2026. The growth has not eliminated the problem; it has changed the design space of what an agent can do without explicit retrieval.
How Ajeris handles this
Ajeris's agent loop bounds each turn with explicit context budgets sized to the channel. A 90-second voice call uses a tighter budget than an asynchronous WhatsApp thread. When context approaches the model's limit, the agent archives older turns into memory and continues; the user does not see the seam.
For most users, context-window management is invisible. The agent just remembers the relevant bits of the conversation and reaches for long-term memory when needed.
