Agentic Memory
Persistent notes or state an agent writes and reads across steps or sessions so it doesn't have to re-derive the same information.
e.g. A running to-do list the agent updates as it completes steps.Context Engineering
Curating the token budget: compaction, just-in-time retrieval, agentic memory.
Compaction
Summarizing or discarding older context to free up space in the context window as a conversation or task grows long.
e.g. Replacing 50 old messages with a 3-sentence summary.Context Engineering
Deliberately curating exactly what information enters a model's limited context window, and in what order, to get better and cheaper results.
e.g. Trimming irrelevant chat history before a new turn.Context Rot
The tendency for model performance to degrade as irrelevant or stale information accumulates in a long context window.
e.g. An agent losing track of the original goal after 100 tool calls.Just-in-Time Retrieval
Fetching information only at the moment it is needed, rather than pre-loading everything upfront, to keep context lean.
e.g. Looking up a file's contents only when a task actually references it.Token Budget
The portion of the context window deliberately allocated to a given kind of content (instructions, history, retrieved docs) so the most important information isn't crowded out.
e.g. Capping retrieved documents to 2,000 tokens so instructions always fit.