Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text.
AgentSoftware built around a language model that gathers information with tools and decides, step by step, what to do next.Append-onlyA store where writers only add new entries and never change old ones; the current state is computed from the entries.ArtifactA file produced by the work and kept for later use: a test report, a generated diagram, a migration script.Assigned confidenceThe confidence on a memory, set by the write policy from its source type, not taken from the model.At-least-onceA guarantee that every step runs at least once, at the price that some steps run more than once.AttemptOne try of a run by one worker process. A resume after a crash is the next attempt of the same run.attempt_numberThe number of one process-level try inside a run: 1, 2, ...Audit trailThe record that lets someone later reconstruct who did what, when and why.AWS Step FunctionsA managed AWS service that runs workflows defined as state machines in JSON, keeps each execution's state and history between states, and can pause a state until an outside system returns a task token.CheckpointSaved execution state that allows a workflow to resume. It answers 'where were we?'.CheckpointerThe part of a system that writes a checkpoint after each step and reads the newest one on resume.CompactionReplacing the older part of a conversation with a summary so that it fits a size limit.CompensationA new action that corrects the effect of an earlier one that cannot be rolled back.Conflict resolutionThe rule that decides what happens when two stored memories contradict each other.ConsentA person's agreement to one specific use of their data, which they can take back.ConsistencyWhether every reader sees the same state, and a state that obeys the rules.ContaminationA memory from one scope (a broker, a submission, a person) reaching a reader in another.Context windowAll the text a model can use in one call: the request plus the response it is writing.Conversation historyThe full, ordered list of messages of a conversation, usually sent back to the model on each call.correlation_idThe id that follows one business request through every system it touches.CrashThe process that runs the work ends without warning: a restart, a kill, a deploy, a power cut.DBOSAn open-source library that records the inputs of each workflow and the output of each step in a system database, so an interrupted workflow resumes from the last completed step with no separate server.DeduplicationRecognizing a repeat of a request or message that was already handled, and not acting on it again.DeterminismThe property that the same input always produces the same output, with no hidden inputs such as the clock or randomness.Disaster recoveryRecovering after the loss of a whole store, disk or site, from backups, within targets set in advance.Document storeA store that keeps whole records, usually JSON, and fetches them by id or by fields inside them.Dual writeCode that writes to two separate systems in sequence and hopes both succeed.Durable executionA way of running a workflow in which every finished step is recorded, so that a run whose process dies continues from the last recorded step instead of starting over.Durable factOne statement kept on purpose for later runs, with where it came from, how far it may be trusted and when it expires.DurableInformation is durable when it survives the end of the process and a restart of the machine.EmbeddingA list of numbers, produced by a model from a text, such that texts with similar meaning get nearby lists.End-state equivalenceThe check that a crashed-and-resumed run ends in the same state as the same run with no crash.Episodic memoryA dated record of something that finished, with the circumstances it happened under.ErasureDeleting a person's data when they ask, including every copy derived from it.Event logThe ordered, append-only list of what happened in a run, from which state can be rebuilt.Event sourcingA design in which the stored record is the sequence of events, and the current state is computed by folding them in order.Evidence idThe id of one stored piece of evidence that a reason or a memory points to.Exactly-onceAn effect that happens once even though the step that causes it may run several times.ExtractionPicking, from what a run saw and did, the statements that might be worth keeping for later runs.File checkpointA saved copy of files taken before an agent edits them, so that a person can undo the edits.ForgettingRemoving or downgrading memories deliberately, by expiry, decay in ranking, supersession or deletion.Hybrid searchRunning a keyword search and a vector search for the same question and merging the two ranked lists.Idempotency keyA value that names one intended effect; the store keeps it with the result and returns that result when the same key arrives again.IdempotentAn operation is idempotent when doing it twice leaves the same result as doing it once.In-process stateValues that live only in the variables of a running program and disappear when it stops.Inbox tableA table on the consumer's side holding the id of every message it accepted, so that a repeat is ignored.InterruptA planned stop in which the run saves its state and waits, with no process running, for an outside input.IsolationHow far one transaction is kept from seeing another's partial or concurrent work.JournalA file where the agent records what is done, what is next and what failed, so the next session can continue.JSON columnA column that holds a structured value whose fields can differ from row to row.Key-value storeA store that maps one key to one value and offers little else, very fast.LangGraph persistenceA library for building agents as graphs, whose persistence layer has two parts: a checkpointer that saves the graph state of a thread, and a store that keeps data across threads.Long-term memoryInformation kept after a task ends so that later tasks can use it.Lost updateTwo writers read the same version, each writes back its copy, and the second write erases the first.Mem0A memory service and open-source library that takes messages, uses a model to extract facts from them, stores those facts with a scope such as a user or an agent, and returns them by search.Memory correctnessThe property that a memory store holds what it should and returns what it should, to the right reader, at the right time.Memory fileA file the coding agent writes itself so that later sessions know something it learned.Memory poisoningGetting false or hostile content written into a memory store so that later runs read it as trusted.MemoryInformation retained for future use. It answers 'what do we know?'.Message relayA separate process that reads unsent outbox rows, sends each one, and marks it sent.Object storageA store for large files, each kept under a name, built for durability and low cost rather than queries.OpenTelemetryAn open-source framework and toolkit for creating and exporting telemetry (traces, metrics and logs) to a backend of your choice; it is not itself a backend that stores them.Optimistic lockingEach record has a version number; a write names the version it read and is rejected if the record has moved on.OverrideAn underwriter's decision that differs from what the agent recommended, recorded with a reason.OwnershipThe rule that each field of a shared record has exactly one writer, and everyone else only reads it.PersistenceWriting information to a store that outlives the process, so that it is durable.Personal dataAny information about a person who can be identified, such as a name with a phone number.pgvectorAn open-source PostgreSQL extension that adds a vector column type, distance operators and approximate nearest-neighbour indexes, so embeddings live in the same database as the rows they describe.Policy fixtureThe invented set of underwriting rules the example applies: UW-GUIDE version 2026.2, effective 2026-07-01.Preference memoryHow one person or role wants to be served, kept until that person changes it.Procedural memoryKnowledge of how to do the work: the steps, the rules, the instructions.Project instructionsA file in the repository, written by people and reviewed like code, that tells the coding agent how work is done here.ProvenanceThe record of where a memory came from: its source type, the evidence it rests on and the run that produced it.Race conditionA bug whose result depends on the order in which two actors happen to run.RankingPutting the candidate memories in order by a score built from a few signals.RecencyA signal that is high for a new memory and falls with age.RecommendationWhat the agent proposes for a submission: approve, decline, request_info or escalate, with reasons, cited rules and a statement that a person must decide.Recovery pointThe moment the restored data comes from; everything after it is lost. Its target, the RPO, is set by the organization.Recovery testA test that crashes the worker at a chosen point, lets the run resume, and checks the result against a run that never crashed.Relational tableA table of rows with fixed, typed columns that a database finds, filters and joins by their values.RelevanceA measure of how well a memory matches the question being asked.Replay boundaryThe line between steps that may be run again during replay and steps whose recorded result must be used instead.ReplayRe-deriving a run's state by reading its recorded events again, without redoing what they did.ResumeContinuing the same run from its last checkpoint, in a new attempt, without redoing the finished steps.RetentionA schedule, set in advance per kind of data, of how long it may be kept, enforced by a job.RetrievalKeeping everything in a store and, at question time, fetching only the passages that match the question.Retry policyThe rules for trying a failed call again within the same step: how many tries and how long to wait between them.run_idThe id of one execution of the workflow for a submission, from its start to a final state.SagaA sequence of steps across several systems in which each step has a compensation to run if a later step fails.SalienceA weight set when a memory is written that says how much it matters, independent of any question.ScopeThe owner of a memory and the readers allowed to see it: the insurer, one broker, one submission or one person.SecretA password, key or token: a value that lets whoever holds it act as someone.Semantic memoryA general statement the agent may rely on later, not tied to one event, stored with its source.Session resumeReopening an earlier conversation from the transcript saved on the machine.SessionOne continuous interaction that an application holds together, such as one chat or one run of a command-line tool.Shared memoryData that more than one agent reads and writes, such as one shared record per submission.Short-term memoryInformation kept only for the current task or conversation. The word does not say which, so ask: short until what ends?Side effectAnything a step changes outside the run's own state: a provider call, a stored row, a sent message.Source of truthFor one field, the single record that wins when two records disagree about it.span_idThe id of one unit of work inside a trace, here one step. 16 hexadecimal characters.Stale factA memory that was valid when it was written and is no longer valid when it is read.State machineA set of named states and the steps that move from one to the next.State migrationChanging saved run state written by old code into the shape new code expects, in storage or when it is loaded.StateThe values a run needs to continue: which step it is at and what it has gathered so far.state_versionThe number of a checkpoint inside one run: 1, 2, 3, counting across attempts.StatelessA stateless service keeps nothing between requests; each request carries everything needed to answer it.StepOne named unit of work in a workflow, such as get_roof_assessment.SubmissionOne broker's quote request, turned by intake from an email into one JSON record.submission_idThe id of one quote request: one broker, one building, one request.Summary driftThe loss that compounds when a summary is summarized again together with new material, round after round.SummaryA shorter text written from a conversation at one moment, kept instead of or next to it.SupersessionReplacing a memory with a newer version about the same scope and subject; the old one stays for audit but is skipped by reads.task_idThe id of one unit of work delegated to an agent or worker inside a run.TemporalA workflow platform in which a service records every event of a workflow execution in an append-only history, and workers rebuild the workflow's state by replaying that history after a failure.Tenant separationThe guarantee that a tenant's data is never read or written on behalf of another tenant, enforced in code.TenantA customer of a shared system whose data must never be read or written on behalf of another customer.The Claude memory toolA tool in the Claude API with which the model asks to view, create, edit, rename and delete files under /memories, while your application carries out each command against storage it controls.thread_idThe id under which a line of work and its checkpoints are filed; several runs can share one thread.TombstoneThe row that remains after a deletion: the memory id, the time and the reason, without the text.Top-kThe number of highest-ranked memories a read returns.trace_idThe id of a trace: the record of what one process did, for people who diagnose. 32 hexadecimal characters.TransactionA group of writes that a database applies all together or not at all.Transactional outboxA pattern that stores an outgoing message as a row in an outbox table, in the same database transaction as the change it announces; a relay sends it later.TTLTime to live: how long a memory stays valid; the store records it as expires_at.Underwriter cockpitThe web application where underwriters see recommendations waiting for them and record decisions.UnderwriterThe person at the insurer who decides whether to accept a risk and on what terms.Underwriting agentThe agent that reads a submission, gathers evidence, applies the rules and writes a recommendation for an underwriter.Vector storeA store or index that keeps an embedding per record and finds the records nearest to a query's embedding.Versioned stateSaved run state that carries the version of its own shape and of the code that wrote it.W3C Trace ContextA W3C Recommendation that defines the traceparent and tracestate HTTP headers, so the trace id of a piece of work travels from one process to the next.WorkerThe process that executes a run: it loads the newest checkpoint, runs the next steps and saves a checkpoint after each.WorkflowA fixed sequence of named steps that turns an input into a result.Working memoryWhat an agent holds while it does one task: the state of the run and what is in the model's input for the current call.Write policyThe rules, in code, that decide whether a proposed memory is stored, with which confidence, scope and expiry, or refused.Write precisionOf the memories the system stored, the share a reviewer agrees should have been stored.