Why does your agent forget the context it needs?
Your agent helped a user perfectly yesterday. Today it greets them like a stranger. That gap is not a bug in your prompt. It is a memory problem.
The problem is structure, not size
Standard RAG pipelines break when enterprises use them for long-term, multi-session agent deployments, as VentureBeat reports. Cramming raw conversation history into the context window makes memory fragmented. Computerworld notes this causes duplicate information and slower retrieval as knowledge grows. The failure is architectural: history is not a memory system.
Organize memory, then retrieve selectively
Two research teams attack the problem the same way. xMemory, from King's College London and The Alan Turing Institute, organizes conversations into a searchable hierarchy of semantic themes. VentureBeat reports it improves answer quality and long-range reasoning while cutting inference costs. Microsoft's Memora takes a parallel path.
- xMemory: hierarchical semantic themes instead of raw history
- Memora: abstraction plus cue-based retrieval, not full conversation replay
The other bet: make context windows huge
EverMind's MSA architecture attacks the problem from the opposite direction. Its press release claims end-to-end long-term memory scaled to 100 million tokens, with strong results on long-context QA benchmarks. This is a vendor press release, not independent coverage. Treat the claims with caution.
What this does not solve
None of these systems solves what to forget, only how to store and find things. Forgetting wrong facts, resolving contradictions between old and new information, and privacy retention rules remain open problems. Microsoft positions Memora as a research project. MSA results come from the vendor itself. None of this is a shipped product you can buy today.
If your agent needed to recall a decision from three months ago, where would that fact live today?
Which of your agent's memories would create legal risk if retained too long?
List the top five facts your agent must recall across sessions, and note next to each whether you would trust a vendor-reported benchmark
Sources
VentureBeat: xMemory cuts token costs in AI agents
Recommended for you
- AI Automation
What breaks when you hand business workflows to AI agents?
AI agents that act, not just chat, multiply what can go wrong. Every permission you grant widens the blast radius of a single mistake.
- AI Automation
When should you automate a job instead of assisting it?
Most leaders automate the wrong tasks. The real decision is which work needs human judgment and which just needs speed.
- AI Automation
When should you build on an AI API instead of a platform?
Meta's enterprise platform is a strategy, not a priced product. Here is how to decide between an API you control and a platform you inherit.
- AI Automation
What breaks when your AI tests and monitoring never talk?
Your pre-launch evals and production monitoring live in separate worlds. That disconnect is why regressions reach customers before they reach your dashboard.
Enjoyed this article?
Subscribe for new articles. No spam. Unsubscribe anytime.