THE CORE AIM OF VOCABULARY MASTERY · AI AGENT MEMORY · CAPTURE → STORE → RETRIEVE → CHECK → UPDATE
What does it mean when an AI agent “remembers” something? AI agent memory vocabulary describes how an AI-enabled application maintains useful information across steps or conversations. It covers context window, session history, working state, long-term memory, episodic memory, semantic memory, retrieval, summarisation, provenance, retention and staleness. These terms are crucial because a stored note is not necessarily true, relevant, current or authorised to reuse.
The core aim of vocabulary mastery here is continuity with correction. Learners should explain what a memory record represents, how it entered the system, who can access it, when it should be recalled, how it is checked against current evidence and what should happen when the record is wrong or no longer needed. Good memory supports continuity; poorly controlled memory magnifies yesterday’s mistake.
This is a specialist owner within the eduKateSG Vocabulary Hub and Core Aim of Vocabulary Mastery learning hub. Read AI Agents Vocabulary for the overall goal–tool–observation loop, AI Agent Orchestration Vocabulary for the flow of jobs between agents, and AI Data Governance Vocabulary for the broader rules on handling data. This page concentrates on the life cycle of remembered context.
Central proposition: An AI memory should be treated as a traceable, revisable information record—not as an unquestionable fact or a licence to keep every detail forever.
The 60-second agent memory router
- Locate: current prompt, context window, message history and working state.
- Classify: short-term, episodic, semantic, procedural or task memory.
- Write: extract a useful fact, retain its source, timestamp and scope.
- Retrieve: search, rank, filter and bring relevant records into working context.
- Check: verify freshness, consent, access rights and contradictory evidence.
- Repair: update, correct, expire, restrict or remove material as appropriate.
AI agent memory glossary
| Term | Meaning | A practical question |
|---|---|---|
| Context window | The information a model can process within a particular interaction. | What is the model seeing right now? |
| Conversation history | Previous messages retained for continuity. | Which past turns are available? |
| Working state | Current task variables, intermediate results and progress. | What has this workflow already established? |
| Session memory | Continuity of information within or across a particular interaction session. | What should carry into the next turn? |
| Long-term memory | Stored information intended to remain available beyond one session. | What should persist for later use? |
| Episodic memory | Records of past events or experiences. | What happened previously, and when? |
| Semantic memory | Stored concepts, facts or relationships. | What knowledge can be reused? |
| Procedural memory | Stored instructions or known ways of performing tasks. | How should this task normally be carried out? |
| Memory extraction | Selecting candidates worth storing from a larger interaction. | What detail merits retention? |
| Memory retrieval | Finding relevant stored information when it is needed. | Which records actually help this request? |
| Embedding | A numerical representation useful for similarity-based retrieval. | How can semantically related notes be found? |
| Provenance | Information about a record’s source and transformations. | Where did this memory come from? |
| Freshness | How recently information was confirmed or updated. | Might the information now be stale? |
| Retention period | The intended duration for keeping information. | How long is continued storage justified? |
| Memory isolation | Separating records by user, organisation or permitted scope. | Whose information may be accessed? |
| Correction or deletion | Updating or removing records when appropriate. | How does an error stop recurring? |
Context is not the same as long-term memory
A model works with information available in its current context. That may include the latest user message, system guidance, retrieved documents and selected previous exchanges. The context window describes the bounded space for that information. When old text is outside it, the model may not see it unless software retrieves or summarises the earlier material.
Long-term memory is a storage design: notes, records or structured facts persist outside the immediate model turn. An application might add selected memories back into the current context when relevant. Therefore, saying “the model remembered” often means “the application saved something and supplied it again”. A user should not assume the model has perfect access to every earlier sentence.
The four useful kinds of agent memory
1. Working or task memory
A workflow can keep the current job title, completed subtasks, draft version, review status and outstanding actions. This is not the same as a permanent user preference. If the job ends, much of the state may be archived briefly for audit or discarded according to its requirements.
2. Episodic memory
An episode records an event: for example, “During the October session, the student confused evaporation with boiling.” The time and context matter because that observation may no longer describe the student after practice. Episodic notes become misleading when a temporary difficulty is treated as a permanent characteristic.
3. Semantic memory
Semantic records capture stable concepts or relationships, such as how a fraction can be represented as a division or ratio. Some facts are broadly reusable; others depend on a curriculum year, source version or school context. Even a factual-looking memory needs enough provenance to determine whether it applies to a new task.
4. Procedural memory
A procedural record describes how to carry out a task: review a claim, check an official source, prepare examples, confirm reading level and only then publish. It is closer to a reusable method than to personal information. Procedures must still be versioned when workflows, policies or available tools change.
A worked example: a Primary 6 Science learning assistant
Imagine a student working through a series of science explanations. In the first lesson, they repeatedly confuse heat with temperature. A useful memory might record the specific misconception, evidence from the student’s answers, the lesson date and the next practice question. It should not casually infer a permanent trait such as “this student is weak at science”.
- Observe: note the exact error in an answer, rather than relying on a vague impression.
- Store selectively: capture the misconception and the teaching step that was helpful; omit unrelated personal details.
- Attach provenance: identify the lesson and context that produced the note.
- Retrieve: when the learner next studies thermal concepts, surface that particular misconception.
- Reassess: ask a fresh question to see whether the difficulty remains.
- Update: replace an outdated “needs help” note with a more accurate current learning state when evidence changes.
This illustrates learning continuity without freezing the student in the past. The memory is a hypothesis to test, not a label to repeat. It can help a tutor reconnect earlier knowledge, but the latest student performance should guide today’s lesson.
Storage, retrieval and use: three different operations
| Operation | What it does | Failure to watch |
|---|---|---|
| Storage | Writes a selected note or record into an authorised store. | Saving a guess as confirmed fact. |
| Indexing | Organises records so they can be found later. | Losing source identity or version. |
| Retrieval | Finds records relevant to the current task. | Surfacing unrelated or wrong-user data. |
| Ranking | Orders records by estimated usefulness. | Preferring a plausible but outdated note. |
| Context insertion | Supplies chosen records to the model during a run. | Treating memory text as higher-priority instruction. |
| Verification | Checks the record against current evidence or user input. | Repeating stale information without challenge. |
| Correction | Changes or removes a record according to valid instructions. | Leaving old copies to override the correction. |
Embeddings and semantic search do not guarantee truth
A memory store may use embeddings to locate text with related meaning even when the words differ. A query about exam anxiety might retrieve a note about stressful revision, for example. That can improve discoverability. But similarity is not credibility: an embedding does not prove that a note is accurate, belongs to the right user or remains relevant. Correct retrieval is one stage; evidence checking is another.
The special danger of stale and contradictory memory
Suppose a student’s timetable changed in September, but a saved note from June still lists their old consultation day. If the agent trusts the old note over an updated calendar, it may give the wrong advice. Strong memory systems preserve dates and sources, compare new statements with existing records and favour current authoritative information. Contradictions should be visible rather than silently blended into an invented compromise.
A useful distinction is recorded at versus valid as of. A note may have been stored yesterday while describing information that was already outdated. When the subject can change, the memory needs a review or expiration rule. This is particularly important for fees, schedules, exam formats, addresses, access permissions and live availability.
Memory boundaries: privacy, authorisation and retention
An application should know which user or organisation a stored record belongs to and which requests legitimately need it. Student learning notes, parent contact details and internal editorial procedures deserve different scopes of access. Memories useful to one learner should not automatically become visible to another. Pseudonymisation or access restrictions can reduce exposure where appropriate, but they do not remove the need for justified collection and retention.
Avoid storing a sensitive disclosure merely because it appeared in a conversation. Keep only information needed for the intended purpose, explain relevant controls to users, honour applicable deletion and retention rules, and avoid turning third-party text into a durable instruction. An uploaded worksheet can provide a clue about a lesson; it should not be able to rewrite an agent’s operating permissions.
Memory summarisation and forgetting
Long histories are often compressed into shorter summaries. Summaries save context space, but they can lose important conditions or the user’s exact wording. If a note says “prefers Saturday lessons when available”, compressing it to “always wants Saturday” changes the meaning. Good summaries preserve negations, exceptions, uncertainty and timestamps. A correction workflow must also account for derived notes or caches that could continue spreading the outdated version.
Design a practical memory record
| Field | Example | Reason |
|---|---|---|
| Subject or scope | Learner session: Science Unit 4 | Prevents cross-user mixing. |
| Claim | Confused boiling and evaporation in two questions | Records an observable result. |
| Source | Practice responses from 14 October | Makes the claim traceable. |
| Confidence or qualification | Observed in this lesson only | Avoids permanent labels. |
| Review trigger | Re-test on the next thermal concepts exercise | Turns the memory into a learning action. |
| Access | Authorised tutor and permitted learner support tools | Restricts unnecessary disclosure. |
| Expiry or retention | Subject to the learning programme’s retention policy | Prompts appropriate disposition. |
A vocabulary task for learners and builders
- Write an example of working state and an example of long-term memory for the same study session.
- Classify these notes as episodic, semantic or procedural: “Today the student made three fraction errors”; “Equivalent fractions name the same value”; “Check prerequisites before giving harder problems.”
- Add source, date, scope and review trigger to a memory that currently says only “student struggles with algebra”.
- Explain why a retrieved memory should not outrank new evidence from a student’s current answer.
- Design one correction path and one deletion or expiry path for a note that becomes invalid.
Common agent memory mistakes
- Keeping everything: more memory can increase privacy exposure and retrieval noise.
- Promoting guesses to facts: distinguish observed evidence from interpretation.
- Ignoring when a fact was valid: review time-sensitive information.
- Mixing user scopes: isolate records and enforce access permissions.
- Trusting semantic similarity as evidence: retrieval relevance is not verification.
- Compressing away qualifications: preserve “might”, “only if” and “no longer”.
- Failing to propagate corrections: inspect summaries and derived stores when a source record changes.
Frequently asked questions
Does every AI chatbot have long-term memory?
No. Persistent memory depends on the application’s design and settings. An AI model’s immediate context should not be confused with an external long-term memory store.
Is memory the same thing as retrieval-augmented generation?
They overlap in technique but differ in purpose. Retrieval-augmented generation may fetch documents or knowledge; agent memory focuses on preserving and reusing relevant state or records across work.
Can memory be wrong?
Yes. A memory can reflect a mistaken inference, an incomplete note, stale circumstances or the wrong user. It should be correctable and reviewable.
Why not save every conversation forever?
Indefinite retention increases storage, privacy, security and confusion costs. Saving should follow a defined purpose and appropriate retention rules.
What is the strongest sign of useful memory?
It helps the learner or worker reconnect with relevant prior information while staying open to correction by more recent evidence.
References and further learning
- OpenAI Agents SDK — Sessions (conversation continuity across agent runs).
- OpenAI Agents SDK — Overview (agents, sessions and tool workflows).
- AI Data Governance Vocabulary (authority, access and data stewardship).
- AI Source Verification Vocabulary (verifying retained claims and their sources).
- AI Agent Orchestration Vocabulary (task state, checkpoints and handoffs).
- immutable eduKateSG learning reference (understanding-led teaching reference).
The AI agent memory vocabulary standard
Mastery means being able to name what is stored, tell working context from durable memory, retrieve the right information for the right person, challenge stale claims, and correct or discard records responsibly. The purpose of memory is better continuity and better decisions—not an archive that can never be questioned.