Kulbir Minhas, AI Initiative. Published 5 November 2026. This piece comes from the AI Initiative learning library: each cited claim was checked against its source (abstract or excerpt) before use. One independent reviewer checked the wording against those claims, and a second, from another model family, confirmed that each reference resolves to the cited paper.
Attribution. Cortex is work at Fifth9. It is used here as the design context because it is already described, with that attribution, on aiinitiative.co.uk.
Why we are asking
Cortex is an agent memory service built at Fifth9. We are designing a memory router for it around two ideas, and both are still design choices to test. The first is to distil at write time: store a compact, checked form of what happened instead of everything. The second is to recall cheaply first: try the inexpensive lookup before the expensive one.
Below is what the research we have checked says, including where it should slow us down. Every cited claim was checked against its source in our learning library (at abstract or excerpt level, not a full re-derivation) before it was used. Preprints are marked, and claims still being checked are left out.
What supports the design
Cognitive architectures for language agents describe memory as an explicit agent component alongside the model (preprint) [1]. MemGPT shows a practical form: tiered, operating-system-style virtual context management lets a model use memory beyond its context window (preprint) [2].
Verbal reflections can improve later attempts without weight updates. Reflexion agents write verbal reflections into an episodic memory buffer and improve on later trials without updating model weights; the paper reports 91% pass@1 on HumanEval [3]. Reflexion motivates our interest in testing what we store; our distillation step remains a hypothesis.
Model routing and cascading for cost-quality trade-offs is an established research direction (survey, preprint) [4]. That survey covers model inference, not memory lookup. Applying the same idea to recall is our design choice.
What should slow us down
A pilot found no statistically reliable cross-model consistency benefit from canonicalising stored memory, after multiple-comparison correction. It used 176 synthetic memory objects and three model families (preprint) [5]. Distilling at write time is a related step, so we will test it rather than assume it helps.
Updates to agent scaffolds and memory can introduce compounding errors (survey, preprint) [6]. We therefore plan to check rewritten memories for mistakes.
Among GateMem's evaluated baselines, none achieved strong task utility, robust access control and reliable agent-facing forgetting together, and the authors note that behavioural forgetting is not verified erasure (preprint) [7]. A survey of machine unlearning also identifies evidence gaps on whether removal persists across external state and later updates in systems with retrieval, memory and tools (preprint) [8].
What we will do
The architectural work supports treating memory as an explicit component [1], and model-routing research motivates testing cost-aware recall [4]. None of this establishes that our distillation approach improves results, and updates to memory can introduce compounding errors [6].
So we will build the router alongside a held-back comparison against storing everything, with a check on each rewritten memory. We will test deletion behaviour and persistence across external state and later updates, keeping behavioural forgetting distinct from verified erasure. We will report which of these held up.
References
- Sumers, T. R. et al. (2023). Cognitive architectures for language agents. Preprint, arXiv:2309.02427. https://arxiv.org/abs/2309.02427
- Packer, C. et al. (2023). MemGPT: towards LLMs as operating systems. Preprint, arXiv:2310.08560. https://arxiv.org/abs/2310.08560
- Shinn, N. et al. (2023). Reflexion: language agents with verbal reinforcement learning. arXiv:2303.11366. https://arxiv.org/abs/2303.11366
- Dynamic model routing and cascading for efficient LLM inference: a survey (2026). Preprint. https://doi.org/10.48550/arxiv.2603.04445
- Memory canonicalization: a framework and benchmark for cross-model drift in persistent LLM memory (2026). Preprint, arXiv:2610.05124. https://arxiv.org/abs/2610.05124
- Self-improvements in modern agentic systems: a survey (2026). Preprint, arXiv:2607.13104. https://arxiv.org/abs/2607.13104
- GateMem: benchmarking memory governance in multi-principal shared-memory agents (2026). Preprint, arXiv:2606.18829. https://arxiv.org/abs/2606.18829
- Xu et al. (2026). Machine unlearning for large language models: foundations, advances, and agentic extensions. Preprint, arXiv:2609.30909. https://arxiv.org/abs/2609.30909