Forget Forgetting

A core limitation of most AI models is their amnesia. Interacting with one is like talking to a colleague who starts every conversation from zero. Your preferences, your project history, and your past corrections all vanish the moment the chat window closes.

As I detailed in my recent article (here), a new class of AI memory systems known as "memory-as-infrastructure" has now emerged in an attempt to address this. Systems like OpenMemory and Mem0 are building a persistent, structured memory layer that can be bolted onto any AI model.

I am certain that memory layers as we are describing here will become pervasive and ubiquitous. The operational advantages are clear and substantial.

And if I'm correct, then the window for shaping how organisations adopt this technology is narrowing.

What AI Memory Infrastructure Actually Does

These AI memory systems promise far more than session persistence. They extract patterns from millions of interactions, maintaining not just data but relationships between data. When configured, they track which factors led to successful outcomes, which communication patterns precede specific decisions, which correlations deserve computational attention.

The breakthrough, as vendors describe it, is multi-dimensional. Memory layers can traverse relationship chains ("Who is Emma's teammate's manager?"), resolve contradictions when new information supersedes old, and maintain temporal validity ("What were my preferences last quarter?"). They achieve this through hybrid architectures: vector databases for semantic search, graph structures for relationship mapping, and temporal indexing for historical queries.

The performance metrics are compelling. Mem0 reports 40% token reduction through intelligent summarisation. OpenNote documented 65% faster case resolution. Enterprises see 80% cost savings compared to feeding full conversation history into each prompt. The systems maintain context across weeks or months whilst keeping latency under 200 milliseconds.

These operational benefits will drive adoption. But before memory infrastructure becomes embedded in every enterprise system, two considerations deserve examination:

Consideration One: Governance.

Memory providers don't just store; they categorise, prioritise, and discard. The frameworks they use to determine significance won't emerge from your organisation's specific context but from their pre-built architectures. We've already seen this homogenisation with LLMs. I wrote about the gravity of the generic pulling every company toward the same linguistic mean. Memory layers threaten to do the same with organisational knowledge.

Will your proprietary processes become their training data? Will your organisational intelligence feed their product improvement?

The economics suggest a pattern: either organisational intelligence trains their systems, or you pay substantially more for memory isolation.

We lack governance frameworks for either scenario. Yet adoption seems likely to follow the LLM trajectory: organisations pursuing immediate efficiency gains whilst governance questions remain unresolved.

Consideration Two: Political economy.

Mem0 advertises installation in "just three lines of code", frictionless adoption that creates profound dependencies. Memory infrastructure could become more strategically important than the LLM itself. Foundation models may commoditise whilst memory layers become the lock-in. The LLM processes; the memory layer determines what persists.

Smart money recognises this dynamic. Hence Olivier Pomel (Datadog) and Thomas Dohmke (GitHub) investing in Mem0. Whoever establishes the memory standard won't need to own the models. They'll control something potentially more valuable: the encoded institutional knowledge of every organisation using their infrastructure.

The existing tech giants, already under regulatory scrutiny and constrained by current architectures, may struggle to capture this market. This leaves space for new entrants to define the category.

Why Biological Memory Is Impossible to Replicate

The institutional and political implications are significant. But there's a more fundamental question that precedes both examinations: can these systems actually work?

If machine memory must function like biological memory, this is impossible.

Yet impossibility hasn't constrained the industry before. LLMs perform reasoning without understanding. AI memory systems will follow the same pattern. But mimicking memory requires something more difficult than mimicking thought: it requires mimicking forgetting.

Forgetting isn't a feature you can engineer. It's an emergent property of constrained biological systems, refined through evolutionary pressure where errors meant death. We can observe its effects but can't explain its mechanisms. Now we're attempting to code what evolution built through extinction.

The implementation reveals the impossibility. Forgetting cannot be programmed because programming requires specification. Autonomous vehicles illustrate this perfectly. Their emergency responses, however sophisticated, are predetermined logic trees. Protect driver or pedestrian, minimise casualties or prioritise youth: these aren't decisions the vehicle makes but rules it executes. The choice happened months before, in code review, not milliseconds before impact.

The same with AI memory systems. When they "forget," they execute deletion rules. Temporal decay, relevance scoring, conflict resolution. All predetermined. The system doesn't forget; it deletes according to specifications.

The machine presents the appearance of emergent properties like deciding or forgetting. It remains programmed behaviour.

Current LLMs demonstrate this perfectly. Recent versions display their "reasoning" process before delivering answers. This isn't reasoning. Like the autonomous vehicle's non-decision, like the AI memory system's non-forgetting, it's pattern-matching that produces reasoning-shaped text. The "thought process" you see is the same statistical assembly as the final output, just formatted to look like deliberation.

This is not a revelation to those working on memory: the industry recognises genuine forgetting is impossible, so instead they're building sophisticated deletion rules. They know the difference; we need to know too.

How the Industry Is Building Around the Impossibility

The technical literature tells a fascinating story. When engineers realised their systems couldn't forget, they didn't give up. They got creative.

They studied how human memory fades. The psychologist Ebbinghaus discovered that we retain only 40% of new information after twenty minutes without reinforcement. So they built "temporal decay" into their systems. Memories artificially losing relevance over time according to the same curve. Except where Ebbinghaus described what happens, they prescribe what should happen. A developer decides the decay rate. Fast for transient data, slow for core patterns.

They studied how we prioritise. Humans somehow know what matters, though we can't explain how. So they built "relevance-based pruning." Every memory scored on multiple factors. How recent, how frequently accessed, how often reinforced. When storage fills, lowest scores get deleted first. It looks like judgment. It's actually a formula.

They studied how we handle contradictions. When new information conflicts with what we remember, something subtle happens in human consciousness. Their solution? Have an LLM evaluate both versions and pick one. They call this "conflict resolution." It's pattern matching against training data, selecting the statistically probable version.

Most revealing of all are the "Ethical Calibration Frameworks." These encode philosophical schools as parameters in decision trees. Utilitarian, deontological, virtue ethics. When the system must choose what to delete, it runs moral philosophy as computation.

The future memory titans are building what can be built: sophisticated deletion orchestration that performs memory well enough for operational purposes.

What Happens When These Systems Reach Your Organisation

It won't arrive as "AI memory infrastructure." It will arrive as the next update to your existing tools. Microsoft will add memory to Copilot. Salesforce will enhance Einstein. Your CRM vendor will announce persistent context as a breakthrough feature. The pitch will be compelling: finally, AI that doesn't make you repeat yourself.

The demonstrations will showcase "intelligent forgetting." Watch as the system knows to discard outdated project details whilst preserving core client preferences. See how it "forgets" irrelevant patterns whilst "remembering" crucial insights. The word "forgetting" will feature prominently. It sounds sophisticated. It suggests judgment, wisdom, discretion.

Behind the demonstrations: temporal decay set to 30 days for project data, 365 days for client data. Relevance scores weighted 40% recency, 30% frequency, 30% utility. Conflict resolution defaulting to newest information unless tagged otherwise. Parameters. Rules. Formulas.

The Gap Between Marketing and Mechanism

When systems claim to "forget" intelligently, they're executing predetermined rules. Someone configured the decay rates. Someone set the relevance weights. Someone chose which philosophical framework governs deletion. Every forgetting decision traces back to a parameter set months before in a configuration file.

The efficiency gains arrive regardless. Faster retrieval, lower costs, better scaling. These systems solve genuine operational problems.

Perhaps that's sufficient. Perhaps the distinction between forgetting and deletion matters less than the operational benefits. Perhaps knowing that Ebbinghaus curves are being applied artificially rather than emerging naturally is academic.

But the impossibility remains. These systems cannot forget because forgetting isn't an operation you perform. It's what emerges from biological constraint, metabolic cost, evolutionary pressure. You cannot engineer what must evolve.

The market will adopt them anyway. Just as LLMs succeeded without genuine reasoning, AI memory systems will succeed without genuine forgetting. The performance will be convincing enough. The benefit, real enough. The impossibility, irrelevant enough.

Until it isn't.

Further Reading and Research

A set of sources that map the emerging architecture of AI memory, from biological analogues to production systems, regulatory pressures, and performance benchmarks. Each reference links to a distinct strand of the field: fast episodic storage, temporal reasoning, unlearning, privacy, and the stability–plasticity tensions that shape long-term behaviour.

Core Architecture and Benchmarks

Chhikara et al. (2024) Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. Shows how two-phase extraction pipelines and hybrid vector–graph stores cut latency and token costs while improving retrieval accuracy. https://arxiv.org/abs/2504.19413

Rasmussen et al. (2025) Zep: A Temporal Knowledge Graph Architecture for Agent Memory. Demonstrates temporal graph construction, bi-temporal modelling and multi-session reasoning gains. https://arxiv.org/abs/2501.13956

Wu et al. (2024) LongMemEval: A Comprehensive Benchmark for Assessing Long-Term Conversational Memory in LLMs. Establishes high-stress memory benchmarks across million-token histories and multi-session reasoning. https://arxiv.org/abs/2410.10813

Maharana et al. (2024) Evaluating Very Long-Term Conversational Memory of LLM Agents. Focuses on 300-turn, multi-session conversations and retrieval precision across extended dialogue. https://arxiv.org/abs/2402.17753

Forgetting, Unlearning and Data Rights

Chen et al. (2022) Graph Unlearning.

A technical foundation for parameter-level removal in graph neural networks, linking deletion requests to model behaviour. https://dl.acm.org/doi/10.1145/3548606.3559352

EU (2017) GDPR Article 17 – Right to Erasure. Defines legal obligations that memory architectures must reconcile with persistent model parameters. https://gdpr-info.eu/art-17-gdpr/

HeyData Team (2025) The Right to Be Forgotten and AI Models Explained Simply. Clarifies why erasure becomes ambiguous once data is embedded in model weights. https://heydata.eu/en/magazine/delete-please-what-the-right-to-be-forgotten-means-for-ai-models

Hine et al. (2024) Supporting Trustworthy AI Through Machine Unlearning. Frames unlearning as a compliance and trust mechanism rather than a technical afterthought. https://link.springer.com/article/10.1007/s11948-024-00500-5

Biological and Cognitive Grounding

D Kline. (2025) Human-Like Forgetting Curves in Deep Neural Networks. Links empirically observed decay functions to configurable retention curves in artificial memory. https://arxiv.org/abs/2506.12034 Enterprise Deployment and Case Studies

Mem0 Team (2025) OpenNote Case Study. 40% token reduction and multi-session continuity without infrastructure changes. https://mem0.ai/blog/how-opennote-scaled-personalized-visual-learning-with-mem0-while-reducing-token-costs-by-40

Mem0 Team (2025) RevisionDojo Case Study. Shows how persistent context eliminates repetitive tutoring loops and increases personalisation. https://mem0.ai/blog/how-revisiondojo-enhanced-personalized-learning-with-mem0

Governance, Local-First Approaches and Alternative Stacks

Getzep Team (2025) Is Mem0 Really SOTA in Agent Memory? Highlights benchmark selection risk and performance variance across contexts. https://blog.getzep.com/lies-damn-lies-statistics-is-mem0-really-sota-in-agent-memory/

Joshi (2025) Bringing Memory to AI Agents. A practical framework for decay, pruning, utility scoring and thresholding. https://prateekjoshi.substack.com/p/bringing-memory-to-ai-agents