"Memory" is the most overpromised word in agent marketing. What's sold as memory is usually just a longer context window — which is not memory, it's a bigger short-term buffer with the same amnesia, just delayed. Real agent memory is an engineered system, and it's simpler than the hype suggests.
What works: structured, written state
What works: structured, written state. After each significant action, the agent updates a compact structured record — what was decided, what's pending, key facts learned. This isn't conversation history; it's a database the agent maintains about its own task. It survives context compaction, it's queryable, and it forces the agent to distinguish "happened" from "assumed." We implement it as a JSON schema the agent is required to update, validated on every write.
What also works: retrieval over past sessions
What also works: retrieval over past sessions. Store summaries of completed tasks (not full transcripts — summaries) in a vector store, and let the agent pull relevant precedents: "how did we handle a case like this before?" This is RAG applied to the agent's own experience, and it's genuinely useful for recurring workflows like support triage or ops tasks.
What doesn't work: dumping entire histories
What doesn't work: dumping entire histories into context and hoping the model "remembers." Long contexts degrade reasoning, inflate cost, and bury the important signal under noise. Also dubious: letting the model freely rewrite its own memory without schema or validation — that's how you get confident false memories propagating across sessions.
The unsexy truth: less memory, better organized
The unsexy truth: most agents need less memory, better organized. A task-scoped state object plus searchable summaries covers the real use cases. Everything else is a research project wearing a product costume.
We build production agents with explicit memory architectures — state schemas, validated writes, retrievable experience — because an agent that can't remember what it did can't be trusted with anything important.
Field note: the index card
Field note: Our support-triage agent's state schema fits on an index card: `issue_summary`, `tried[]`, `customer_facts{}` (plan tier, prior tickets, sentiment), and `recommended_action` with a confidence flag. Every session ends by writing a 3-line summary to the experience store. When a similar ticket arrives weeks later, the agent retrieves the precedent: "last time a tier-2 customer hit this error, the fix was X and it took one reply." Resolution time on repeat issue classes fell by half. The memory isn't magic — it's just the agent's own notes, structured, validated, and searchable. Boring, reliable, compounding. Start with the index card, not the research paper — your agent will thank you.
Related service Workflow automation
