Enterprise discussions about AI often begin with model capability.
Which model reasons best? Which one has the largest context window? Which one performs best on a benchmark?
But once agents begin doing real work, another limitation quickly appears.
They forget.
A conventional agent can appear remarkably intelligent during a single session and surprisingly ignorant the next morning. Previous decisions disappear. User preferences need to be explained again. Project history has to be reconstructed. Lessons learned during one workflow rarely improve the next one.
This is less like hiring a digital colleague and more like hiring an extremely capable contractor who develops amnesia every night.
That is why persistent memory is becoming one of the most important areas in agent architecture.
Oracle introduced an AI Agent Memory initiative in 2026 around exactly this problem: without memory, agents effectively begin each interaction from scratch. Microsoft Research has meanwhile been exploring several approaches to long-term agent memory, including PlugMem, Memora, MemGym and human-inspired architectures.
The interesting point is not simply that agents should store more history.
In fact, storing everything creates another problem.
A transcript is not a memory
One of the most useful insights emerging from current research is the difference between history and memory.
Raw conversation history records what happened.
Useful memory captures what matters.
Microsoft’s PlugMem research argues that dumping long interaction histories back into an agent can overwhelm it with low-value context. The proposed alternative is to transform interactions into structured, reusable knowledge.
That distinction is crucial for enterprise systems.
Consider a project agent that participates in six months of conversations.
The raw record might contain:
- thousands of messages;
- duplicated project descriptions;
- superseded deadlines;
- tentative decisions;
- corrected information;
- social conversation;
- multiple versions of the same plan.
The valuable memory might be:
- the client chose option B;
- legal approved the revised wording;
- launch moved from October to November;
- James owns final approval;
- a previous attempt at the workflow failed because the CRM data was incomplete.
Those are not merely excerpts from a transcript.
They are organisational state.
Memory also needs time
Enterprise truth changes.
A customer changes supplier. A project changes owner. A price changes. A policy is superseded. A deadline moves.
A serious memory system therefore cannot merely ask, “What facts have we stored?”
It also needs to understand:
- what is believed to be true now;
- what was believed to be true previously;
- when something changed;
- what evidence caused the change;
- what was known at a particular point in time.
That becomes particularly important for agents that participate in business decisions.
An agent may need to answer two very different questions:
“What is the current agreed price?”
and:
“What price did the team believe had been agreed when this decision was made?”
Those questions require temporal memory, not just vector search.
Memory compounds enterprise value
Knowledge bases create value when people search them.
Memory systems create value simply by being present while work happens.
Every interaction can improve the future context available to the agent.
A useful correction today reduces the chance of a repeated mistake tomorrow. A decision captured in one workflow can inform another. A user’s working preferences can persist across models and applications. The organisation gradually builds a reusable layer of operational intelligence.
This is where agent memory becomes much more than a convenience feature.
It becomes infrastructure.
The Iridium view
Iridium separates Memory, Knowledge and Tools because they solve different problems.
Knowledge answers:
What approved information can this agent rely on?
Tools answer:
What systems may this agent access or act through?
Memory answers:
What has this organisation, team or user learned through previous activity?
The distinction matters.
Enterprise agents should not depend on one model vendor to remember the organisation. The durable intelligence should live outside the model so it can move with the business.
Models will change.
Institutional memory should not disappear with them.
Sources
- Oracle, Introducing Oracle AI Agent Memory: A Unified Memory Core for Enterprise AI Systems (2026)
- Microsoft Research, PlugMem: Transforming raw agent interactions into reusable knowledge (2026)
- Microsoft Research, Human-Inspired Memory Architecture for LLM Agents (2026)
- Microsoft Research, MemGym: a Long-Horizon Memory Environment for LLM Agents (2026)