What are the types of AI agent memory?

Enterprise example. A procurement agent checking a purchase order holds, in its window, the system instructions, the order itself, the supplier record it just fetched from the ERP, the contract clause a retrieval step returned, and the last few turns with the buyer. When the task ends, that window is gone unless something writes it elsewhere.

Two limits shape it. Size: windows have grown (Anthropic made a 1M-token window generally available on 13 Mar 2026), but every token sent into it is billed, and the Lost in the Middle study (2023) found models use information in the middle of long inputs worse than at the ends. Overflow: tools exist to trim and summarise. Anthropic’s context editing clears old tool results on the server, by default at 100,000 input tokens while keeping the last three tool uses; compaction (Anthropic beta, OpenAI compact threshold) replaces older turns with a summary.

Between turns, working memory is saved outside the model and replayed: a LangGraph checkpointer, an OpenAI Conversations object, AgentCore short-term events (immutable, timestamped, grouped by actorId and sessionId, kept up to 365 days) or Agent Platform Sessions on Google. The context engineering sub-page covers how to spend the window well.

Enterprise example. A field-service agent remembers that a technician covers the night shift, prefers parts from the Hamburg depot and writes reports in German. Payhawk’s Financial Controller Agent, built on Google’s Memory Bank, remembers spending habits and auto-submits small meal expenses; the vendor reports submission time cut by more than 50% (Google Cloud blog, 18 Dec 2025, vendor story).

Product names: AgentCore’s semantic and user preference strategies, Foundry’s user profile memory (retrieve it early in a conversation, Microsoft advises), Memory Bank’s memory profiles (GA 15 Jul 2026), Copilot’s saved memories and the details it infers from chat history.

Controls: scope per user, because a preference is personal data; a TTL at creation (AgentCore minimum 31 days, Foundry store default TTL in the preview) so a stale preference expires; consolidation so a new fact replaces the old one; and a way for the user to see and delete what is stored, which Copilot offers in Settings and which you must build yourself on the developer services.

Enterprise example. An IT service-desk agent resolving a VPN error recalls that for the same error last month it reset the user’s certificate, the user confirmed the fix, and a reboot was needed first. The episode becomes a few-shot example on the next ticket of that kind.

Products: AgentCore’s episodic strategy with reflections (added December 2025; records appear only after AgentCore detects that an episode has finished), Foundry’s chat summary memory and AgentCore’s summary strategy for the lighter form, and Memory Bank’s consolidation across sessions.

Controls: this is the type that leaks. AWS warns that episodic reflections can span multiple actors and recommends guardrails or keeping reflections at the actor level. An episode also contains whatever the user typed, so the PII screen on the write path matters most here, and the erasure runbook must cover reflections as well as events.

Enterprise example. An accounts-payable agent learns that invoices from one supplier always need a purchase-order match before approval, and that a particular approver wants amounts in euros. Microsoft’s Foundry procedural memory stores reusable how-to routines and operating patterns inferred from prior interactions and retrieves them when a recurring task comes up; Microsoft reported a 7 to 14 point absolute gain on the Tau-bench benchmark at Build (2 Jun 2026, vendor-reported). It is in preview.

Other forms: LangChain describes the system prompt as procedural memory refined through reflection, and LangMem offers prompt optimisation; Anthropic’s memory tool lets the model keep notes-to-self files under a /memories path that your application stores, which teams use for learned procedures; AGENTS.md files play the same role for coding agents.

Controls: a procedural memory changes behaviour for every user in its scope, so treat a write to it as a prompt change. Review and version it, then evaluate the agent again before it takes effect. We never let an agent update this type in production without a human approval.