Overview
This series examines how mksglu/context-mode keeps large outputs, durable evidence, and long-running session state outside an agent’s active conversation until needed. It moves from deliberate computation and selective retrieval to separate source and work-state memories, compaction recovery, security controls, and host-specific lifecycle limits. The central thesis is conditional: context savings depend on returning the right derived result or exact evidence, while freshness, arbitrary execution, local persistence, and host capability remain operator responsibilities.
Learning outcomes
- Context-mode reduces context transfer by routing raw inputs to computation, indexing, or session storage; it does not make context costs disappear.
- Sandbox computation suits derived facts, while indexed retrieval preserves exact wording, code, and qualifications when fidelity matters.
- Content memory and work-state memory have different owners: one recalls sources, the other restores objectives, files, decisions, and errors.
- Security and continuity depend on host hooks, policy, execution controls, freshness settings, and staged verification, so adoption claims must stay conditional.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What problem does mksglu/context-mode solve?
mksglu/context-mode addresses context overload by keeping raw data, searchable evidence, and session state outside the model’s active conversation until needed. It returns derived answers, selected passages, or a compact resume guide rather than every raw byte.
When should an agent use computation instead of reading a file?
Use computation when the task needs a count, filter, parse, comparison, or aggregate rather than exact bytes for inspection or editing. Use native reads for exact text, known lines, and edits; computation reduces returned output but not authorization responsibility.
How does context-mode preserve work after compaction?
It captures structured events in hooks, stores them in SessionDB, builds a prioritized resume snapshot before compaction, and injects a compact guide on resume; detailed events remain searchable. Coverage depends on the host hooks and retained events.
Does context-mode guarantee security and context savings on every host?
No. Its tools can travel across hosts, but enforcement and automatic continuity depend on lifecycle APIs; execution retains host filesystem and network authority, and measurements vary by workload and denominator. Operators must verify hooks, policies, storage, freshness, and results on the actual host.