AI Memory Needs More Than a Relevance Check, Preprint Argues

Marvin Danig of Achiral Research sets out the proposal in Jacobian-Causal Memory Control for Persistent Memory in Language Model Systems, submitted to arXiv on November 5, 2026. The work is an architectural specification and mathematical analysis, not a demonstration of a working system’s superiority. It reports no benchmark measurements and releases neither an experimental dataset nor benchmark code.

From matching information to predicting influence

The paper distinguishes two questions: does a stored item match the current request, and what will including it cause the model to do? Conventional retrieval uses signals such as textual similarity, metadata or relationships between records. Danig proposes adding an estimate of the direction in which each authorized memory would shift the model’s behavior.

That estimate is called an “influence signature.” Signatures and observed effects would share a measurement space, which the paper calls model J-space. Despite the name, this need not represent the model’s internal state. It could contain scores from behavioral probes, selected token probabilities or other output measurements.

The name “Jacobian-causal” refers to a mathematical tool for describing how small input changes affect outputs. Where model internals are accessible, derivatives could help estimate a memory’s influence. Where they are not, the paper proposes controlled comparisons, probes or learned approximations. No single estimator is presented as universally suitable.

A requester or policy process would specify the intended behavioral change, prohibited effects and a side-effect budget. The controller would select memories whose predicted influence best fits that specification, accounting for uncertainty and policy risk. Hard access and provenance restrictions would exclude candidates outright, rather than merely lowering their scores.

Checking the prediction

The proposed feedback loop compares matched model runs with and without a memory intervention. Each pair uses the same runtime state and, where possible, the same random seeds or sampling draws. This helps separate the memory’s effect from variation in generated answers. Without shared randomness, uncertainty estimates must account for sampling variation.

Here, “causal” means this controlled intervention contrast—not a claim that ordinary activity logs reveal causation.

If the measured effect differs too much from the prediction, the system could remove context, revoke a memory-tool result, restore a checkpoint, suppress future selection or quarantine the memory. Records of predictions, observations and corrective actions would preserve lineage and help assess future reliability.

But rollback has a hard boundary: it cannot undo consequences that have already left the model. Removing a stale summary cannot recall an email an agent has already sent.

What the mathematics establishes

The paper proves four conditional results. First, if the response changes sufficiently smoothly, the error in a local linear prediction is bounded by a quantity proportional to the intervention’s size squared. This supports small, local approximations—not reuse across arbitrary prompts or model versions.

Second, under an additive approximation with unrestricted mixing weights, the available signatures determine which target effects are reachable. Authorization and other deployment constraints can only restrict that reach further.

Third, if every candidate’s estimated loss differs from its true loss by at most a known amount, selecting the best estimate incurs at most twice that amount in extra loss compared with the best authorized choice.

Fourth, a smoothness bound limits how much two memories’ combined effect can depart from the sum of their separate effects. Strong interactions require direct joint measurement instead.

None of these results establishes that practical signatures are accurate or that the controller beats simpler retrieval.

The tests still to come

The proposed evaluation would compare full JCMC with ordinary retrieval, graph-based recall, policy-filtered retrieval and influence-based ranking without observation. Comparisons would hold model, memory, policy and cost budgets equal, measuring accuracy, intended and unintended effects, rollback recovery, calibration and operating costs.

The proposal also introduces risks. Signatures might disclose private information; probes might miss important behaviors; individually benign memories might interact harmfully. Steering toward a target effect could itself become manipulative unless constrained by user intent and policy. Estimation adds expense, especially through closed-model interfaces, and signatures can become stale.

The paper explicitly allows its central claim to fail. If simpler systems match effect targeting and side-effect control more cheaply, the claim narrows. If observation cannot reliably distinguish beneficial from harmful interventions, or rollback and lineage do not improve recovery, the proposed loop fails its stated tests. For now, JCMC offers a specification to investigate, not a validated solution.