Engineering / research
Plan with bounded context and memory
Analyze working state, memory retrieval, search, replanning, and termination without assuming longer traces or larger memories imply correct reasoning.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 research
Plan with bounded context and memory
Analyze working state, memory retrieval, search, replanning, and termination without assuming longer traces or larger memories imply correct reasoning.
- 01
Working context is a bounded view of versioned task state
Select task-relevant context while preserving goals, constraints, evidence identities, and pending action state rather than assuming a longer transcript is better memory.
32 min - 02
Episodes preserve scoped observations while semantic memory stores derived claims
Separate event-level provenance from compressed generalizations and evaluate retrieval cost without treating self-reflection as verified world knowledge.
33 min - 03
A memory write needs purpose, provenance, and an explicit trust boundary
Choose what to persist without promoting tool-output instructions into authority or allowing repeated and unsupported observations to become durable facts by accident.
34 min - 04
Retrieve for relevance only after checking scope, validity, and revocation
Separate retrieval ranking from evidence eligibility and revalidate source-dependent memories before their use in a current decision.
33 min - 05
A task graph must encode evidence and effect dependencies as well as subgoal names
Make artifact versions, completion evidence, shared-state conflicts, and retry semantics explicit before parallelizing an agent's plan.
34 min - 06
Search spends a finite budget on candidates whose evaluations may be wrong
Account for generated and evaluated branches, pruning, verification, and shared candidate errors without treating more search as guaranteed better reasoning.
34 min - 07
Replanning reconciles observations with preconditions while preserving the original task
Update a versioned plan after actual tool evidence, reconsider dependent work, and reconcile uncertain effects before retrying them.
34 min - 08
A belief state needs an observation model; a planning score is not a probability
Use a small explicit partially observed model to distinguish state uncertainty, Bayesian evidence updates, heuristic rankings, and decision utility.
35 min - 09
Stop on verified task conditions or explicit limits, not the appearance of a finished trace
Define completion evidence, measurable progress, retries, and deadline reserves so an agent can report incomplete work without fabricating success.
34 min - 10
Evaluate memory and planning with matched tasks, actual cost, and failure-specific controls
Separate the effect of stored information, retrieval, search, verification, and extra compute while testing stale, misleading, duplicate, and adversarial evidence paths.
35 min