Agents
Tool use, orchestration, memory, and the engineering of autonomous systems.
Articles
7 PUBLISHEDUnderstanding Agent Procedural Memory: How Should This Task Be Performed?
Procedural memory stores approved, versioned playbooks. Retrieve the active procedure, execute its structured steps, persist the run, and ground the answer in current results.
Understanding Agent Episodic Memory: What Happened Last Time?
Episodic memory preserves complete experiences. Store each issue, action, observation, resolution, and outcome, then retrieve similar resolved cases as evidence.
Understanding Agent Vector Memory: What Knowledge Is Relevant?
Vector memory retrieves durable knowledge by meaning. Embed explicit memories, rank semantic matches, apply filters and thresholds, and keep retrieved content untrusted.
Understanding Agent Key-Value Memory: When Exact Lookup Wins
Key-value memory stores typed, durable facts at stable addresses. Use exact lookup, scoped keys, replacement semantics, and TTL without involving the model in storage.
Understanding Agent Working Memory: What Should Stay in Context?
Working memory is the bounded context for the next model call. Build a policy that drops noise, summarizes older state, protects recent turns, and still guarantees a token budget.
Understanding the ReAct Loop: Is It Just Observe → Think → Act?
ReAct turns tool results into feedback. Build the loop manually, then compare it with the orchestration provided by the OpenAI Agents SDK.
Understanding the Ralph Loop: Is It Really Just a While Loop?
Observe–Think–Act tells an agent how to work. Ralph gives fresh agents repeated chances to continue that work using persistent state.