Ones to watch
Memory, forgetting and reasoning are becoming the next control points for trustworthy agentic AI.
Agentic persistent memory
Why it matters:
Persistent memory allows agents to retain context across tasks and sessions, moving from one-off assistance to continuous working relationships. The governance issue is equally important: retained memory must be scoped, auditable and removable, so systems that remember can also forget.
Evidence:
Mem0’s 2026 State of AI Agent Memory report frames memory as a first-class agent infrastructure layer, with LoCoMo, LongMemEval and BEAM emerging as benchmarks. Vectorize’s
Hindsight adds a product signal: an open-source memory system built around retain, recall and reflect operations, structured time-aware memory, and parallel dense, sparse, graph and temporal retrieval. Vectorize reports 94.6% LongMemEval performance and publishes benchmark materials but the result should still be treated as product-reported unless independently replicated in the reader’s target context. Sources: Mem0, 2026; Vectorize Hindsight, 2026.
Machine unlearning
Why it matters:
Machine unlearning is becoming a practical control for removing the influence of targeted data, behaviours or capabilities without rebuilding a model from scratch. It is relevant to erasure requests, consent withdrawal, safety remediation and contested training data.
Evidence:
SISA, or Sharded, Isolated, Sliced and Aggregated training, reduces the retraining burden for unlearning by isolating the affected shard and reusing cached training states but it is not a general guarantee of exact deletion for modern foundation models. Influence-based methods and attribution can narrow the removal target, while negative fine-tuning can reduce harmful behaviours but is not the same as guaranteed data deletion. MUSE’s evaluation of eight unlearning algorithms on 7B-parameter models shows methods can reduce memorisation while still struggling with privacy leakage, utility preservation and repeated large-scale removal requests. Sources: Bourtoule et al., 2021; Shi et al., MUSE, 2024.
Energy-based models for reasoning
Why it matters:
Energy-based reasoning models score how consistent a partial reasoning trace is with a problem’s constraints, allowing failures to be detected and repaired mid-plan rather than only after a final answer is produced.
Evidence:
Logical Intelligence positions Kona as an energy-based reasoning system and Aleph as an orchestration/proving layer; secondary coverage reports very high PutnamBench performance. This is a useful signal for high-stakes reasoning where correctness, inspection and repairability matter, but the claim should be treated as a vendor or market signal pending broader independent benchmark validation. Source: Logical Intelligence, “Energy-Based Models for AI Reasoning”, 21 January 2026.
Tech Radar
How to read the radar
Readiness indicates how close a technology is to practical enterprise adoption. Impact indicates the expected strategic or operational significance for organisations. Positions may change between editions as evidence, adoption and maturity evolve.
This Tech Radar builds on technologies introduced in the earlier AI Horizon editions, including our last one. It is designed as a live monitoring mechanism rather than a fixed assessment, tracking how selected technologies move across readiness and impact as the market matures. Changes in position reflect developments in technical maturity, enterprise adoption, regulatory relevance, implementation evidence and practical industry use. Each new edition will add emerging technologies to the radar, while existing entries will continue to be monitored to identify whether they are maturing, stabilising or declining in relevance.
Conclusion
The agentic AI landscape for enterprise organisations is transitioning from one of experimentation to one of operation and action. Standardised infrastructure through MCP, convergent frontier model capabilities, mature security frameworks and proven ROI pathways now exist.
The boardroom questions of the moment are no longer whether agents work, it is which organisations have designed their workflows, governance structures and measurement systems to sustain agents at scale and therefore avoid project cancellation.
As we signalled at the start of this edition, the organisations succeeding are embedding agents into defined workflows with measurable outcomes, not attempting enterprise-wide deployments without the necessary checkpoints. They are treating governance and security as essential pillars, not things to be retrofitted. The window for this transition is closing and there is now enough evidence to suggest that the organisations who establish governance and measurement frameworks now will be the ones who compound their advantage going forward.