A context window is the finite sequence of tokens a language model can see in a single call. It acts as short-term working memory: instructions, evidence, and tool results compete for limited space. Token counts determine what fits and often influence latency and cost. section 1.1 section 4.1.1
Treat context as a budget, not uniform RAM. Every token admitted consumes capacity, while excluded information is unavailable to the model. Position matters: different parts of the window do not necessarily receive equal treatment. section 3.1 section 1.1
Keep frequently needed working state close to the model, and retrieve durable history or source material when its value justifies the token cost. Paging information into context changes its availability, not its authority. Provenance, ownership, version, permissions, and invalidation remain separate checks. section 4.1.1
Table 4.1 Dimensions of agent memory
|
Memory category
|
Example content
|
Scope
|
Function
|
Lifetime and access
|
Owner or authority
|
|
Active context
|
Current instructions, evidence, tool results
|
Turn or session
|
Working state
|
Immediate and volatile
|
Runtime plus source owners
|
|
Task or session state
|
Goal, artifacts, receipts, pending decisions
|
Task or session
|
Resumption
|
Durable until task closure
|
Orchestrator or task owner
|
|
Project or user context
|
Conventions, preferences, profiles
|
Project, user, or team
|
Continuity
|
Durable, versioned, retrieved on demand
|
Human or system owner
|
|
Admitted lessons
|
Scoped guidance with admission evidence
|
Workload, team, or organization
|
Governed reuse
|
Versioned, expiring, supersedable
|
Admission owner
|
|
Archival evidence
|
Raw traces, prior versions, audit records
|
Task through organization
|
Audit and later review
|
Archived under retention policy
|
Evidence or records owner
|
Compare active context with longer-lived memory by scope, function, access, and authority. section 4.1.1





































