Skip to content

Systems and records

Agent memory retention tiers: which business facts should agents keep?

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Decide by testing each fact before an agent stores it: who is the authority for it, when it was true, what permission allows keeping it, and what should happen when it changes. Facts that pass become dated, sourced claims in one of five retention tiers, from never stored to durable with history. A fact with no source, owner or exit rule stays out of persistent memory.

Agent run

Agent memory retention tiers

This page covers agent memory retention tiers: which business facts should agents keep.Illustrative

Key takeaways

  • Persistent agent memory should hold claims with a source, a valid time and a permission basis, not loose sentences the model found useful once.
  • Several memory stores keep items indefinitely by default, and some refresh expiry on every read, so retention has to be set per fact class by the team.
  • Five tiers cover most business facts: do not store, session only, cached pointer to a system of record, dated claim with a review date, and durable reference with history.
  • Every tier needs an exit trigger: a source change, a review date, a correction, a withdrawn permission or a deletion request.
  • Superseded facts should usually be invalidated and kept as history for audit, but kept out of answers about the current state.
  • Regulated retention, legal holds and third-party personal data are decisions for a records owner and counsel, not for the agent.

How do you decide which business facts belong in persistent agent memory?

A business fact belongs in persistent agent memory only when four answers are written down before it is stored: the authority that can assert it, the time it was true, the permission that allows keeping it, and the event that ends its life. If any of the four is missing, the fact stays in the current session or is read fresh from the system of record each time.

This framing treats memory as a store of claims rather than a notebook. A claim such as the billing contact for this account is the finance lead says more than its text: it came from a specific record, it became true on a date, someone is allowed to change it, and it can be wrong later. An agent that stores only the sentence loses every one of those properties, and the first time the contact changes the agent repeats an old answer with full confidence.

The practical output is a retention tier for each class of fact. Tiers turn a vague instinct about what the agent should remember into rules an engineer can implement and a records owner can review.

Why a store's default settings are not a retention policy

Memory stores ship with defaults chosen for convenience, and those defaults rarely match how long a business fact stays true. Reading the configuration of the store you use is the fastest way to see the gap.

In the LangGraph store, the default time-to-live for new items is none, meaning no expiration. When a TTL is set, the timer refreshes on reads as well as writes by default, so a fact the agent retrieves often never ages out, which is the opposite of what a stale but popular fact needs. Expired items can still be returned by reads unless the omit_expired option is turned on.

Amazon Bedrock AgentCore Memory separates raw short-term events, kept for a configurable period of up to 365 days, from extracted long-term records. AWS documents that long-term records do not expire by default; an optional per-strategy TTL has a 31-day minimum, and expired records are removed by a background process rather than at the exact moment.

Other tools leave the whole decision to you. Anthropic documents its memory tool as client-side: the model asks for file operations and the application's own handler decides where the /memories files live and how long they stay. The Mem0 README describes its current extraction as add-only, where memories accumulate and nothing is overwritten, and says facts the agent itself generates are stored with equal weight to facts from users. None of these is wrong, but each moves the retention decision onto the team deploying the agent.

The retention tier decision table

Use the table to assign every class of fact an agent might write to one tier, then implement the expiry and exit rule for that tier in the store. Assign classes, not individual facts, so the decision is made once and reviewed by the records owner. The examples name common business systems for illustration.

The retention tier decision table
TierWhat belongs hereExamples from business systemsExpiry and refresh ruleExit trigger
T0 Do not storeSecrets, payment and identity numbers, sensitive records with no stated purpose, and values that change faster than the agent can learn themAPI keys pasted into a Slack thread, a card number read out on a call, a live account balance in NetSuiteNever written to memory; redact from any session summary before it is savedNot applicable; a write attempt is logged and blocked
T1 Session onlyWorking context for the task in hand, including what a person said they will do nextThe Zendesk ticket being worked, a draft reply, the caller's description of the problemDiscarded when the task closes; a summary is promoted only if each fact passes the four testsTask closed, handed off or abandoned
T2 Cached pointerFacts owned by a system of record that are slow to fetch; memory keeps the pointer and a short-lived copyCurrent plan tier in the billing system, open order status in the ERP, this week's on-call rota in PagerDutyShort fixed expiry measured in minutes to days, never refreshed by reads; the agent re-reads the source before actingExpiry reached or a change event from the source
T3 Dated claim with review dateFacts asserted by an authority that stay true for months but do change, stored with valid-from and review-by datesBilling contact in Salesforce, a customer's stated channel preference, a negotiated service window in the contract recordNo silent expiry; on the review date the agent re-verifies against the source or stops using the factSource change, review date passed without confirmation, correction by a person
T4 Durable reference with historyStable facts and approved decisions that later work depends on, kept with every versionLegal entity name, a published refund policy version, an approved exception recorded with its approver in JiraKept until withdrawn; superseded versions are invalidated and kept as history, not deletedNew version approved, policy withdrawn, deletion request or end of the retention period

Which fields must a fact carry before an agent stores it?

A fact is ready for tiers T2 to T4 only when it carries enough metadata to be checked, corrected and removed later. Without these fields, the tier is a label with nothing to enforce it.

At a glance

Which fields must a fact carry before an agent stores it

Two times matter, and they are easy to confuse. Valid time is when the fact was true in the business; recorded time is when the agent learned it. Graphiti's edge model makes the split explicit with a valid_at field for when a fact became true, an invalid_at field for when it stopped being true, and an expired_at field for when the system invalidated it. A memory record that only stores the write timestamp cannot tell a change that happened last week from a backfill of something that changed a year ago.

  • Fact text and the subject it is about, such as an account, a contact or a policy, using the source system's identifier.
  • Source system and record identifier, plus the field or message the fact was read from.
  • Authority: the role or system allowed to assert or change this fact, for example the account owner in the CRM or the finance team for billing terms.
  • Valid from and valid to, kept separate from the recorded time.
  • Permission basis: the purpose the fact may be used for and any restriction from contract or privacy notice.
  • Tier, expiry rule and review-by date.
  • Correction history: prior values, who corrected them, when and why.
  • Deletion behavior: whether removal must reach summaries, embeddings and backups derived from the fact.
  • Conflict status: whether another claim of equal authority currently disagrees.

Update, stop using or forget: which trigger calls for which action?

Update when an authority changes the value, stop using when the fact can no longer be confirmed, and forget only when permission ends or someone asks for deletion. These are three different actions, and collapsing them into a single delete either loses the audit trail or keeps data the business should not hold.

Graphiti's documentation describes the update path well: when information changes, the old fact is invalidated rather than deleted, and each fact traces back to the raw episode that produced it. That pattern keeps the answer to what was true in March available while keeping it out of answers about today.

Update, stop using or forget: which trigger calls for which action?
TriggerAgent actionKeep history?Who confirms
Source record changes the valueWrite the new claim, set valid-to on the old oneYes, as supersededNo one if the source is the authority; the record owner if it is not
Review date passes without confirmationStop using the fact in answers and actions; flag it for re-verificationYes, marked unconfirmedThe fact's authority or the record owner
A person corrects the agentRecord the correction with name, time and reason; supersede the old claimYes, with the correctionThe correcting person must hold authority for that fact
Permission withdrawn or purpose endedRemove the fact and anything derived from it from retrievalOnly a minimal log that removal happenedPrivacy or records owner
Deletion request from the person the fact is aboutDelete the fact, its summaries and embeddings; follow the backup policyNo content kept; keep the request recordPrivacy owner, with counsel where the law is unclear
Two claims of equal authority disagreeUse neither for action; answer with both and the datesYes, both keptThe record owner resolves

Worked example (illustrative): six facts from one support escalation

Consider an illustrative escalation: a customer calls about failed invoice exports, the agent reads the Zendesk ticket and the Salesforce account, and the call ends with a promised follow-up. Six candidate facts come out of the session. The table shows the tier each one lands in and why. Durations are illustrative and should come from the team's own policy.

Worked example (illustrative): six facts from one support escalation
Candidate factSource and authorityTierRule applied
Caller read out a card number to prove identityCall audio; no business purpose for memoryT0Blocked and redacted from the call summary
The exports started failing on MondayCaller's account; unverifiedT1Useful for this ticket only; promote only if engineering confirms it in the incident record
Account is on the annual Enterprise planBilling system is the authorityT2Pointer kept with a one-day copy; re-read before any plan-related promise
Billing contact is the finance leadSalesforce contact role set by the account ownerT3Valid-from date stored; review in 90 days or on a CRM change event
Customer prefers email over phone follow-upsStated by the customer on the callT3Stored with the call reference as its source; review in 180 days; customer can change it anytime
Credit approved under refund policy version 4Approval recorded by a support managerT4Kept with the policy version and approver; superseded only by a new approval

The exception: facts whose value is the history itself

Some facts should never expire in the usual sense, because their value is that they record what the business believed or decided at a point in time. Approved exceptions, policy versions in force on a date and decision records with their approvers fall here.

The rule for these is to keep the history and change how it is retrieved. An agent asked whether a customer is eligible for a credit today should see only the current policy version, while an auditor asking why a credit was granted in March should see the version in force then and the person who approved it. Storing both under one tier with valid-time filters handles both questions; deleting the old version on update breaks the second one.

Legal holds are the other exception in the opposite direction. A hold can require keeping records that the tier table would otherwise expire, and it overrides the table until the hold is lifted by counsel.

Where the tier table stops and a records owner decides

The table assigns fact classes to tiers; it does not settle how long the law or a contract requires a business to keep or delete a record. Those decisions belong to the records owner and counsel, and the tier table should point to their retention schedule rather than repeat it.

Personal data is the common case. In the EU, the GDPR's storage limitation principle in Article 5(1)(e) says personal data should be kept in identifiable form no longer than necessary for the purposes it is processed for, and Recital 39 asks controllers to set time limits for erasure or periodic review. A T4 durable tier holding customer contact details needs that analysis; a T4 tier holding published policy versions does not.

Stop and ask for specialist review when memory spans several client companies in one store, when the agent writes inferences about people such as sentiment or intent, when facts come from regulated records like health or financial data, or when a deletion must reach model weights or derived datasets. Agent-generated facts deserve particular care: a tool that stores them with the same weight as facts from a system of record gives a guess the same standing as an authority.

Buying memory histories that keep their corrections

Teams building or testing memory systems need business records where facts change, get corrected and get withdrawn on real dates. Event logs, CRM field histories and ticket audit trails carry that signal, but only when the change history, authority and timestamps survive export.

SourceX does not stock memory-history datasets; it finds them on request among U.S. businesses, with the buyer stating the scope, volume, format, timing and licensing terms the memory work needs. It runs rights review, contracting and preparation with the supplying company, and nothing is delivered until that company has approved the scope and terms and signed.

Frequently asked questions

What are agent memory retention tiers?

They are classes that set how long an agent may keep a type of fact and what ends its life. A common set runs from facts that are never stored, through session-only context and short-lived cached pointers, to dated claims with review dates and durable references kept with their history. Each class gets an expiry rule and an exit trigger.

Is a time-to-live setting enough to keep agent memory accurate?

No. A TTL measures age since a write or a read, not whether the fact is still true. In some stores reads refresh the timer, so a frequently used fact can outlive its truth indefinitely. Pair TTLs with source change events and review dates.

Should an agent remember facts it inferred on its own?

Only as clearly labeled inferences with the evidence that led to them, and usually only for the session. An inference has no authority behind it until a person or system of record confirms it. Storing it with the same weight as a sourced fact lets a guess shape later answers.

Does deleting a memory remove it everywhere?

Not automatically. Summaries, embeddings, caches, logs and backups made from the fact can survive the deletion of the original record. A deletion procedure should list each derived store and say how removal reaches it, and the privacy owner should confirm it for personal data.

How long should short-term agent memory be kept?

As short as the task needs. Keep working context until the task closes or is handed off, then discard it or promote individual facts that pass the authority, time, permission and exit tests. Product limits, such as a maximum event retention period, are ceilings rather than recommendations.

Sources

  • The LangGraph store's TTL defaults to no expiration, refreshes on reads by default, and expired items can be returned unless omit_expired is enabled. Source
  • AgentCore Memory keeps raw short-term events for a configurable period of up to 365 days. Source
  • AgentCore long-term memory records do not expire by default; an optional TTL has a 31-day minimum and expired records are removed in the background. Source
  • Anthropic's memory tool is client-side, so the application's handler controls where memory files are stored and how long they persist. Source
  • Mem0 describes add-only extraction in which memories accumulate and are not overwritten, with agent-generated facts stored at equal weight. Source
  • Graphiti edges record valid_at, invalid_at and expired_at, separating when a fact was true from when the system invalidated it. Source
  • Graphiti invalidates changed facts instead of deleting them and traces each fact to the source episodes that produced it. Source
  • GDPR Article 5(1)(e) limits keeping identifiable personal data to what the purpose needs, and Recital 39 asks for erasure time limits or periodic review. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies