Machinery outside the model.
AI memory, as a field of research and engineering practice, names the techniques that let a language-model-based agent persist, organize, and selectively recall information across interactions. A base model retains nothing between calls: each call sees only the tokens placed in front of it. Whatever an agent appears to remember, a prior instruction, a past mistake, a fact established three sessions earlier, has to be captured, kept somewhere, and handed back in as input on a later call by a system built around the model.
That system is the subject of this field. It ranges from pasting a short summary back into the next prompt to maintaining a structured store queried across sessions. A survey covering work on the topic from 2022 through early 2026 treats it as a distinct sub-discipline with its own operational definition, its own taxonomy, and its own open problems, rather than as a single mechanism (arXiv:2603.07670).
Three coupled operations.
The survey formalizes agent memory as three operations tightly coupled to an agent's perception and action, rather than a single store-and-fetch step.
| Operation | What happens |
|---|---|
| write | Something from the current interaction is captured for later use: an outcome, an error, a piece of context the agent produced or observed. |
| manage | What was captured is organized: compressed, summarized, indexed, scored, or discarded. This step is where systems diverge most, since it decides what survives to be read back. |
| read | On a later call, some subset of what was managed is selected and returned to the model as input, in whatever form the system stores it in. |
A position, not a category.
The same survey places a given memory system at a point along three independent axes, rather than sorting it into one named category.
| Axis | What it categorizes |
|---|---|
| temporal scope | How long retained information persists: local to a single turn, bound to one session, or durable across sessions. |
| representational substrate | The form information is stored in: raw context tokens, vector embeddings, structured or symbolic records, or parameters updated by training. |
| control policy | What decides which information is written, kept, evicted, or recalled: a fixed rule authored in advance, or a policy learned from data. |
Five answers to the manage step.
Current approaches group into five mechanism families, each a different answer to what the manage step of the loop does with what was written.
| Mechanism | Description |
|---|---|
| context-resident compression | History is summarized or compressed so it fits inside the model's context window, rather than stored outside it. |
| retrieval-augmented stores | Information is held in an external index, most often a vector store, and a subset is fetched by similarity to the current query. |
| reflective self-improvement | An agent examines its own past behavior and writes higher-level notes or rules, consulted on later calls rather than the raw history itself. |
| hierarchical virtual context | Memory is organized into tiers and paged in and out of the active context, similar in shape to virtual memory in an operating system. |
| policy-learned management | What to write, keep, or evict is decided by a component trained for the purpose, rather than by a fixed heuristic. |
Where the field has gaps.
The same survey names five open problems the mechanism families above do not yet resolve: continual consolidation of memory over long-running use, causally grounded retrieval that distinguishes what actually influenced an outcome from what merely resembles it, trustworthy reflection that does not compound an agent's own errors, learned forgetting that discards information deliberately rather than only by running out of space, and multimodal, embodied memory that extends beyond text. None of the five mechanism families is presented as resolving all five gaps at once.
One position in the taxonomy.
Substrate memory, and three of its component questions, are placed within the vocabulary set out above rather than treated as a separate field.
Common questions.
Is AI memory the same as retrieval-augmented generation (RAG)?
No. Retrieval-augmented stores are one of five mechanism families a recent survey identifies within the field; RAG names one common instance of that family. AI memory as a field also covers context-resident compression, reflective self-improvement, hierarchical virtual context, and policy-learned management.
Is there a single agreed definition of AI memory?
Several surveys converge on a similar operational description, a write-manage-read loop coupled to an agent's perception and action, but differ on which axes and mechanism families they name. The field is actively being taxonomized rather than settled.
Does AI memory require training or fine-tuning a model?
Not in general. Of the five mechanism families named above, only policy-learned management involves learning a component; the other four can operate over an unmodified model.
Is "AI memory" an established term?
Yes, in the sense used on this page: a body of research and engineering practice with its own surveys, taxonomies, and benchmarks, distinct from any single mechanism or vendor. This page follows that established usage rather than a term coined by the Vinculum project.