In brief
- The model does not automatically have persistent memory; the application supplies stored context again.
- Working state, session history, long-term memory and authoritative business data are different kinds of storage.
- Memories can be wrong, stale or overly broad and therefore need provenance, validity and deletion rules.
- Agent memory must not become an uncontrolled replacement for CRM, ERP or other systems of record.
What kinds of memory are there?
During a task, an agent keeps intermediate results and the current process state. A session history preserves the dialogue for one case. Long-term memory stores selected preferences or facts across sessions. The agent can also retrieve authoritative information from business systems when needed.
These forms should be treated separately. A temporary calculation needs different retention and access rules from a customer preference; a confirmed contract status still belongs in the system responsible for that record.
| Storage type | Purpose and typical duration | Reliable source | Main risk and deletion rule |
|---|---|---|---|
| Working memory | Intermediate steps for one task; minutes to hours | Current run and verified tool results | Errors propagate; delete when the run is complete |
| Session memory | Dialogue and state for one case; until case closure | Messages, events and confirmed case inputs | Wrong association; close or delete under the retention rule |
| Long-term memory | Selected preferences or facts; across sessions | Confirmed information with provenance and validity | Stale or sensitive memory; provide expiry and correction |
| System of record | Authoritative business state; domain retention applies | CRM, ERP or accountable business database | Memory conflicts with source; retrieve current data instead of copying |
When does memory help, and what can go wrong?
Memory avoids repetition, supports longer tasks and can respect relevant preferences. It also creates new failure modes: information may be inferred incorrectly from a conversation, assigned to the wrong person or reused after it has changed.
Invisible context is especially problematic. If users cannot see which memory influenced an answer, they have little chance to correct a false assumption. Relevant memories should therefore be visible, editable and traceable to their origin.
- Use continuity only where it demonstrably improves the task
- Do not treat a model-generated summary as a confirmed fact
- Resolve conflicts in favour of current authoritative system data
- Confirm uncertain or sensitive memories before reuse
Which data and access rules does agent memory need?
Persistent memory is a separate data store and must be inventoried, protected and operated accordingly. Purpose, legal basis, location, retention, deletion and access rights need to be clarified independently of the model being used.
- Step 1
Limit the purpose
Store only information that is genuinely needed for a defined future task.
- Step 2
Set the scope
Separate memory reliably by person, organisation, case and role.
- Step 3
Manage the lifecycle
Define expiry, updates, correction and automatic deletion.
- Step 4
Trace access
Authenticate and log who reads, writes and deletes memories.
How should memory be designed and tested?
Start without long-term memory and add it only for explicitly named information types. Define who may create a memory, how it is confirmed and when the system must query current data instead.
- Test changed, conflicting and deliberately false memories
- Probe tenant, role and user separation for data leakage
- Compare answers with and without memory to measure real value
- Give users a clear way to correct and delete stored memories
Example: a procurement agent with case memory
Within one procurement case, an agent remembers which offers have arrived and which questions remain open. It does not infer preferred supplier terms from conversations; it retrieves them from the approved supplier master. The temporary working state is deleted after completion, while only the decision recorded in the business system remains authoritative.
What to remember
Treat agent memory as a controlled data store with purpose, provenance, validity and deletion. Authoritative business data remains in the systems that own it.
Sources and further reading
These primary sources provide further detail on definitions, technical foundations or responsible use.
Content reviewed