> ## Documentation Index > Fetch the complete documentation index at: https://docs.apollospace.ai/llms.txt > Use this file to discover all available pages before exploring further. # What is AI agent memory? > The category, defined independently of any one vendor: the three kinds of memory a production AI agent needs, why a fresh context window every session isn't memory, and how Apollo Space's agents implement it. ## Definition **AI agent memory** is the mechanism that lets an agent retain and reuse information across conversations and over time, instead of starting from a blank context window every session. It's what separates an agent that "learns" your organization from one that has to be re-briefed on the same facts every time you open a new chat. Memory is a category with real structure to it — not one undifferentiated blob of "stuff the agent remembers." Production agent architectures generally distinguish three kinds: Discrete facts: dates, preferences, who owns what, what was promised to which customer. The kind of thing you'd write down in a note. Approaches that worked before — how a particular kind of task tends to get done well in this organization, learned from repetition rather than stated once. Insights distilled from many conversations — patterns and conclusions that emerge across dozens of interactions, not traceable to any single one. ## Why a long context window isn't the same thing as memory A model with a very large context window can hold a lot of text in a single conversation, but that's not memory in the sense that matters operationally: it doesn't persist once the conversation ends, it doesn't get reused by a *different* agent working on a related task, and it isn't structured — a wall of raw transcript is not the same as a curated fact, a learned procedure, or a distilled insight. Real agent memory is retained **across** sessions and agents, not just held within one long session. ## How it's typically retrieved The dominant production pattern is **semantic search** over a memory store — an agent asks "is anything relevant to what I'm doing right now," and gets back the facts, procedures, or insights that match, rather than requiring an exact keyword. This is what lets a new agent, or the same agent on a new task, draw on context nobody had to re-explain. ## Apollo Space's implementation Apollo Space's shared memory layer is the **[Company Brain](/en/features/brain)** — organization-scoped and versioned, holding documents, knowledge captures (web snippets with source URL and a note), the org's voice document, and **agent memories**: facts, preferences, and learned patterns accumulated from actual operation rather than hand-written. Every [agent](/en/concepts/agents) that needs context runs a semantic search against the Brain before acting, and what one agent learns becomes available to the others working in the same organization — with the [Digital Twin's](/en/agents/digital-twins) personal memory scoped privately to the individual it represents, never leaking into the org-wide Brain. Isolation matters here as much as retrieval: memory is scoped per organization, so one customer's Brain never surfaces in another's agent context. See [Multi-tenant](/en/trust/multi-tenant) for how that boundary is enforced. ## What memory doesn't do Memory retrieval is not the same as judgment. An agent recalling that a discount was promised last quarter doesn't mean it should apply a new one — the [Agents](/en/concepts/agents) architecture keeps a budget and a trust boundary around what memory-informed actions an agent can take autonomously versus what it has to propose to a human first. ## Next steps Apollo Space's implementation of the memory layer, in full detail. How memory fits alongside persona, tools, and budget. How memory stays isolated between organizations. The category that depends most visibly on memory working well. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.