**DRAFT — pending editorial expansion.** This article is a working draft published as scaffolding for the NINtec content programme. The current version covers the substantive perspective in compressed form; the published version will expand each section to the 2,000+ word depth the topic warrants. Editorial review is required before promotion.
Long-running agents need memory that survives process restarts, scales beyond context windows, and remains queryable for relevant past interactions. Claude's 200K context window is large but not infinite; production agent memory is an architectural pattern, not a model feature.
Durable memory stores
Postgres, Redis, vector databases — the backing store depends on the access pattern. Sequential conversation history fits Postgres; semantic recall fits vector databases; structured agent state fits Redis. Most production memory systems use multiple stores.
Summarisation and compression
When context exceeds the window, summarisation compresses older interactions while preserving salient detail. Multi-tier summarisation (recent verbatim, mid-term summarised, long-term highly compressed) is the durable pattern.
Retrieval-grounded recall
Semantic retrieval over the agent's memory store surfaces relevant past interactions for the current task. The pattern is RAG over conversational memory — same engineering discipline applies (chunking, embedding, retrieval evals).
Process-restart durability
Production agents survive process restarts without losing context. Checkpoint-and-resume patterns, idempotent action invocation, and structured state externalisation are the engineering primitives that deliver this.
Memory architecture is the difference between agents that remain coherent across sessions and agents that lose context with every restart. Production-grade agentic systems include this from architecture phase.