Engineering classification guide

Types of AI agent memory

Working, episodic, semantic, and procedural memory provide a useful shared vocabulary for AI agent design. No one list settles every design question because records can differ by functional role, representation form, lifecycle operation, retention horizon, and authorized scope. Treat the cognitive terms as design analogies, then classify each retained object across the dimensions that matter to its workload.

Memory guide topics
On this page

Classification method

Why agent memory type lists disagree

Agent memory type lists disagree because they flatten different dimensions into one hierarchy. Some labels name an item's job. Others name form, duration, operation, or access boundary. Labels become useful when their axes are explicit. The method below is guide-created, not a field standard, and must be adapted to the workload.

Illustration

This guide's classification method

representation form × functional role × lifecycle operation × retention horizon × scope

One record can occupy several positions at the same time. The five dimensions are independent rather than a serial lifecycle.

Dimensions in reading order

  • Representation form names how information is carried.
  • Functional role names the job the information serves.
  • Lifecycle operation names what happens to the record.
  • Retention horizon names which boundary the record survives.
  • Scope names where authenticated authority permits access.

External research CoALA contributes proposed functional modules T05. Hu and colleagues contribute a forms, functions, and dynamics frame T06. Zhang and colleagues survey sources, forms, and operations T07. No cited source proposes this exact combination.

Category errors to reject

  • A vector index is an access mechanism, not a functional role.
  • Short retention does not make a cache working memory.
  • An entity view does not raise source authority.
  • A procedure does not grant execution permission.
  • A delayed intention needs durable workflow state.

Functional roles

What functional roles can agent memory serve?

Functional roles name the job retained information performs. Working, episodic, semantic, and procedural memory provide a core vocabulary. Factual, experiential, prospective, and entity roles cover other workload objects. These cognitive terms are design analogies, not human equivalence. One record can serve several roles, but no label establishes truth or authority.

External research Baddeley and Hitch ground working memory T01. Tulving distinguishes episodic and semantic memory T02. Squire reviews interacting human memory systems T03. The software terms are analogies, not a biological model.

Core roles

Working memory
Facts, intermediate results, and control state used in current work. Active use, not duration, defines the role.
Episodic memory
Events or trajectories with time, source, participants, and outcomes. Later change does not erase the historical event.
Semantic memory
Reusable knowledge not bound to one event. Derivations should link to supporting episodes.
Procedural memory
Methods, skills, code, or reviewed instructions. Retrieval remains separate from execution authority.

External research Peer-reviewed UIST work on Generative Agents uses a memory stream, retrieval, reflection, and planning in its simulation T08. Voyager, a peer-reviewed TMLR paper, uses a retrievable executable skill library in Minecraft T10. Both are specific designs, not a complete taxonomy.

Extended roles

Factual memory
A proposition with source, time, and status. The label does not make it true.
Experiential memory
Cases, strategies, or lessons derived from outcomes. It can overlap other roles.
Prospective memory
A delayed intention tied to a time, event, or state cue. It needs task state and cue detection.
Entity memory
An entity-centered view of attributes and relations. It requires identity resolution beyond a display name.

External research Rummel and Kvavilashvili review prospective memory as memory for delayed intentions T04. That cognitive account does not define an agent scheduler.

External research Peer-reviewed AAAI work on ExpeL retains trajectories and derives insights in evaluated tasks T09. PM-Bench, a COLM paper with a linked arXiv record, evaluates delayed intentions and cues in its published task T11. Neither defines a universal implementation.

Representation forms

What form can an agent memory take?

Representation form names the physical or logical carrier of retained information. It may be an explicit record, model parameter, continuous state, or hybrid of source material and access aids. Form does not determine function, retention, or authority. A form without addressable records needs an inspection and erasure contract.

Explicit or symbolic
Text, fields, events, relations, or executable procedures that can be addressed and inspected.
Parametric
Information encoded in model weights or adapters, with update and deletion boundaries tied to the model process.
Latent or continuous
Hidden state or dense representations whose meaning may require a decoder or downstream task.
Hybrid
A sourced record paired with derived summaries, relations, or retrieval representations.

Separate record and retrieval contracts

Illustration

A record contract names identity, source, time, version, status, scope, authority, and retention. A retrieval contract names queries, filters, ranking, stale behavior, and limits. An embedding is commonly a derived retrieval representation, not the authoritative record. Another design must declare its provenance, update, access, and erasure rules.

Retention and scope

How long and where should memory persist?

Retention horizon says what boundary an item survives, while scope says who may write, retrieve, change, export, or delete it. A short horizon does not make a record working memory, and a broad horizon does not make it shared. Each design must name the authenticated scope and the event that ends retention, which may differ across derived copies.

Retention horizons with retrieval and ending conditions
Horizon What survives Who can retrieve it What ends retention
Inference call Current model input and temporary state The active runtime The call completes
Run Intermediate task state across calls The authorized run Completion, cancellation, or expiry
Session State needed across one interaction session Actors and services in that session Session policy closes it
Cross-session subject Sourced history for a resolved subject Authorized work for that subject Retention policy or scoped erasure
Workspace Team tasks, methods, and reviewed lessons Authorized workspace members and agents Workspace policy or object removal
Shared organization Approved organizational knowledge Authorized organizational roles Withdrawal, replacement, or policy expiry
Archive Historical evidence retained for a declared purpose Restricted archive roles Archive schedule or erasure duty

Scope comes from authenticated authorization. Model arguments, display names, or retrieved identifiers may narrow but cannot establish access.

Short-term and long-term are retention horizons, not functional roles. A short-term item may be episodic or factual. Working state can survive through a durable checkpoint. Living AWS AgentCore documentation uses short-term and long-term as one current vendor vocabulary T12. Those labels are not a neutral standard.

Workload selection

Which memory roles does this agent need?

Choose memory roles by listing the objects later work must recover, not by selecting a storage product first. For each object, decide its source, authority, retrieval trigger, update rule, and erasure rule. The result is a workload contract, not a universal taxonomy. Some agents need a small role set, while others need overlapping views.

Workload objects and candidate functional roles
Workload object Candidate role Selection question
Active task state Working What must survive the next step or retry?
Event history Episodic Which event, time, source, and outcome matter later?
Reusable claims Factual and semantic Which source and validity state support reuse?
Methods Procedural Which version applies, and who may use it?
Learned lessons Experiential and semantic Which outcomes support the derivation?
Delayed intentions Prospective Which cue, owner, task state, and cancellation rule apply?
Entity-centered view Entity plus underlying roles How is identity resolved and each field sourced?

Write the chosen object's source and authority before choosing its retrieval trigger. Then define correction, supersession, expiry, and erasure behavior. Use the architecture guide to assign those contracts to components and the evaluation guide to turn them into cases.

Retrieval-augmented generation, or RAG, retrieves external items for a query and conditions generation on them. It can provide an access path for records classified here, but it does not determine their role, scope, or update semantics. The agent memory and RAG guide owns that comparison.

Synthetic inventory

Incident-response memory inventory

Synthetic classification inventory for INC-204. It explains contracts and reports no measured result.

  1. Active mitigation hypothesis

    Object or claim
    Active mitigation hypothesis
    Source
    Incident commander note and linked telemetry
    Functional role
    Working and factual candidate
    Representation form
    Structured claim with evidence links
    Retention horizon
    Incident
    Scope
    INC-204 response team
    Lifecycle state
    Unconfirmed candidate
    Retrieval trigger
    Mitigation planning or hypothesis review
    Authority
    Operational hypothesis, not service truth
    Retention and erasure check
    Expire at closure or supersede when evidence changes
  2. Failed deployment and rollback attempt

    Object or claim
    Failed deployment and rollback attempt
    Source
    Deployment system and rollback log
    Functional role
    Episodic event
    Representation form
    Explicit records with event time
    Retention horizon
    Service incident history
    Scope
    Checkout operations
    Lifecycle state
    Observed event with later outcome
    Retrieval trigger
    Timeline review or another rollback proposal
    Authority
    Event evidence from operating systems
    Retention and erasure check
    Retain by incident policy and erase covered identifiers
  3. Current service owner

    Object or claim
    Current service owner
    Source
    Live ownership system
    Functional role
    Factual and entity-oriented
    Representation form
    Structured entity projection
    Retention horizon
    Cross-session
    Scope
    Authenticated incident workspace
    Lifecycle state
    Current until rota change
    Retrieval trigger
    Escalation or assignment
    Authority
    Live ownership system
    Retention and erasure check
    Replace on rota change and clear stale projections
  4. Approved rollback runbook RB-17

    Object or claim
    Approved rollback runbook RB-17
    Source
    Versioned runbook registry
    Functional role
    Procedural
    Representation form
    Document with structured steps
    Retention horizon
    Workspace
    Scope
    Authorized Payments SRE roles
    Lifecycle state
    Approved version
    Retrieval trigger
    Rollback consideration
    Authority
    Method guidance, not execution permission
    Retention and erasure check
    Withdraw the version and its retrieval entries
  5. Derived comparison lesson: compare pool saturation before repeating rollback

    Object or claim
    Derived comparison lesson: compare pool saturation before repeating rollback
    Source
    Review linked to supporting incidents
    Functional role
    Experiential and semantic
    Representation form
    Derived claim with source links
    Retention horizon
    Cross-session workspace
    Scope
    Checkout reliability team
    Lifecycle state
    Reviewed derivation
    Retrieval trigger
    Before repeating rollback
    Authority
    Lower than supporting event evidence
    Retention and erasure check
    Recheck when support changes and erase covered derivatives
  6. Delayed certificate check

    Object or claim
    Delayed certificate check
    Source
    Incident action item
    Functional role
    Prospective
    Representation form
    Durable task and named trigger
    Retention horizon
    Until completion, cancellation, or expiry
    Scope
    Checkout deployment workflow
    Lifecycle state
    Pending
    Retrieval trigger
    Next deployment reaches its stated wait point
    Authority
    Intention, not execution permission
    Retention and erasure check
    Scheduler tracks completion and removes expired task copies
  7. Checkout service view

    Object or claim
    Checkout service view
    Source
    Projection over sourced records and live systems
    Functional role
    Entity view joining roles, not a new cognitive store
    Representation form
    Hybrid entity projection
    Retention horizon
    Workspace
    Scope
    Authorized checkout responders
    Lifecycle state
    Derived current view
    Retrieval trigger
    Service page or incident query
    Authority
    Each field retains its own authority
    Retention and erasure check
    Rebuild or erase projections when sources change

The runbook cannot authorize a rollback. The delayed check cannot wake itself. A scheduler or workflow must hold the trigger state, while a trusted host still decides whether a permitted operation may run.

Bounded product mapping

Current ContextDB mapping

Current ContextDB behavior

As checked against repository snapshot d59c1b9 on , the pinned Apache SDK is Available. Direct factual extraction is Alpha. Asynchronous Formation is Hosted Alpha. The SDK and hosted paths implement a narrower set of memory operations than the neutral roles described here. Hosted Formation produces sourced factual candidates. Those facts do not mean ContextDB forms working, episodic, procedural, experiential, prospective, entity, or multimodal memory.

Types FAQ

Questions about agent memory types

Can one agent memory record belong to more than one type?

Yes. A deployment event can be episodic because it records what happened, factual because it states an outcome, and entity-oriented within a service view. Keep each role explicit. Multiple labels do not increase authority. Source, time, status, and scope still determine how later work may use the record.

Is short-term memory the same as working memory?

No. Short-term names a retention horizon, while working memory names active task use and manipulation. A session cache may be short-lived without supporting current work. Working state may also be durably checkpointed across calls. Agent designs use both terms by analogy, so name role and horizon separately.

Is entity memory a separate cognitive type?

Entity memory can be an entity-centered organization or projection, not evidence of a separate software faculty. It can assemble facts, events, procedures, and pending tasks for one resolved entity. The design must preserve each item's source and role. Identity resolution must not rely on a display name alone.

When does an agent need prospective memory?

An agent needs prospective memory when it must retain a delayed intention, then react to a named time, event, or state cue. The record should include task state, trigger, owner, cancellation, and expiry. A scheduler or workflow must detect the cue because storing an intention does not execute it.

Is an embedding itself a memory?

An embedding can carry a representation for similarity retrieval, but it is not inherently authoritative memory. One contract keeps the sourced record, status, version, time, and scope separately, then uses the embedding as an index entry. Another contract must still explain provenance, updates, access, and erasure.

How should procedural memory be scoped?

Scope procedural memory to the actors, agents, environments, versions, and tasks allowed to use the method. Keep execution authority separate. A rollback runbook can explain approved steps without permitting service changes. Supersede withdrawn procedures, and apply erasure or archival rules to copies and retrieval entries.

Source ledger

Primary sources and maturity

The ledger records direct primary URLs, publication maturity, and the boundary of each source. Cognitive research supplies analogical terms. Agent research supplies proposed designs and evaluated examples, not a universal taxonomy.

  1. External research Baddeley and Hitch, Working Memory, 1974. Foundational scholarly chapter. Grounds the human working-memory term, not an agent implementation.
  2. External research Tulving, Episodic and Semantic Memory, 1972. Foundational scholarly chapter. Grounds the human distinction, used here by analogy rather than as software equivalence.
  3. External research Squire, Memory Systems of the Brain, 2004. Peer-reviewed review. Supports interacting human systems, not separate software stores.
  4. External research Rummel and Kvavilashvili, Current Theories of Prospective Memory, 2023. Peer-reviewed review. Grounds delayed intentions but does not define an agent scheduler.
  5. External research Sumers and colleagues, CoALA, 2024. Peer-reviewed TMLR paper. Supports proposed working, episodic, semantic, and procedural modules.
  6. External research Hu and colleagues, Memory in the Age of AI Agents, 2025 (revised 2026). External preprint. Supports a proposed forms, functions, and dynamics frame, not consensus.
  7. External research Zhang and colleagues, ACM TOIS, 2025. Peer-reviewed survey. Supports a survey frame across sources, forms, operations, and limits.
  8. External research Park and colleagues, Generative Agents, 2023. Peer-reviewed UIST paper. Illustrates a memory stream, retrieval, reflection, and planning in its simulation.
  9. External research Zhao and colleagues, ExpeL, 2024. Peer-reviewed AAAI paper. Illustrates retained trajectories and derived insights in its evaluated tasks.
  10. External research Wang and colleagues, Voyager, 2024. Peer-reviewed TMLR paper. Illustrates an executable skill library in the studied Minecraft setting.
  11. External research PM-Bench, 2026. COLM conference paper with linked arXiv record. Supports delayed-intention and cue evaluation in its published task.
  12. External research AWS AgentCore memory types. Living vendor documentation. Illustrates one vendor's labels, not the neutral classification method.

Compare one record with current implementation

Inspect which memory operations exist today and which neutral categories remain outside the current product scope.