Jacobian-Causal Memory Control for AI Memory

Persistent memory changes the effective input distribution and policy context under which a language model is sampled. A model with durable user, project, or organizational memory is no longer answering from a blank prompt. It is being steered by what the system lets it remember.

That makes memory recall a control problem.

A memory can preserve project state, recover institutional context, or help an agent avoid repeating a mistake. It can also shift the frame of a recommendation, make a stale fact feel current, or amplify a pattern the user never meant to reinforce. The relevant object is therefore not merely a retrieved record. It is an intervention whose behavioral effect should be anticipated, bounded, observed, and, when necessary, reversed.

Research question

The research question is:

Under what conditions does governing persistent AI memory by estimated behavioral effect, rather than relevance alone, improve reliability, auditability, and recovery in long-lived language model agents?

The essay and companion preprint treat this as an architecture-and-analysis problem. They do not report benchmark measurements, production thresholds, private corpora, or deployment policy. The contribution is a public evaluation framework for memory recall as an observed inference-time intervention: expected effect, governed selection, post-application measurement, lineage, and recovery.

The central hypothesis is narrow:

For long-lived agents and organizational memory, recall should be evaluated as an observed, lineage-bearing intervention, not as relevance-ranked context alone.

Public abstraction

Jacobian-causal memory control is the public name for a research framework that estimates how a candidate memory is expected to change a model's behavior, compares that expectation with observed behavior, and studies what recovery evidence is needed when the estimate is wrong.

We use model J-space as a public abstraction for the local behavioral-effect space in which those estimates can be compared. The term is intentionally narrower than "latent space." It does not imply direct access to a model's hidden state, and it does not refer to a published production representation, threshold, or estimator. It names the research object: a measurement space in which an intervention can be described by direction, magnitude, uncertainty, and context dependence.

Let $y$ denote a behavioral measurement under a fixed probe protocol $P$, model version $\theta$, decoding policy $d$, and task context $c$. For a memory $m$, a public-level influence estimate can be written as:

$$\Delta_P(m \mid c) = \mathbb{E}[y \mid c, m, \theta, d] - \mathbb{E}[y \mid c, \varnothing, \theta, d],$$

projected into a task-relevant measurement space. This notation is not a disclosure of a production estimator. It states the scientific comparison that any claimed implementation should make auditable.

In cognitive-systems language, this is loosely analogous to activation in cognitive architectures such as ACT-R, where prior knowledge becomes differentially available to cognition. Jacobian-causal control asks a distinct systems question: once those memories become available to a language model, how are they expected to move observable behavior, and how should a system respond if the observed movement diverges from prediction?

Limits of relevance

Most retrieval systems rank stored items by relevance to a query:

$$s_q(m_i) = \operatorname{sim}(f(q), f(m_i)).$$

That score helps nominate candidates. It does not specify the causal role the memory will play once it enters an inference-time pathway.

A pricing note can be relevant to a vendor recommendation while pulling the answer toward cost savings when the intended behavior is compliance caution. An old project summary can match the query while suppressing newer contradictory evidence. A memory can be authorized in isolation and still become risky when combined with the current prompt, tool state, or another recalled memory.

The real-world analogy is a design review. A prior incident report may be highly relevant to the current decision, but relevance is not the whole question. The team also needs to know whether introducing that report will make the discussion more cautious, more conservative than necessary, more focused on reliability, or less sensitive to new evidence. Causal influence is about that shift in reasoning, not merely the presence of a matching document.

The distinction is familiar in other forms of systems engineering: an input can be admissible, relevant, and still destabilizing under a particular operating condition. Persistent memory needs an analogous notion of intervention effect.

Influence signatures

An estimated interventional effect signature is a public-facing abstraction for a directional effect estimate. It asks: if this memory is made available now, which aspects of the model's behavior should change, by how much, and with what uncertainty?

The public requirement is comparability, not disclosure of a particular internal estimator. Predicted and observed effects must be expressed in a shared influence space, with enough provenance to make disagreements auditable across model versions, prompt families, memory states, and measurement protocols.

This deliberately leaves implementation choices out of the public essay. A practical system may use different measurement regimes depending on model access, latency budget, and risk context. Those choices are treated as protected engineering and evaluation work, not as public recipe material.

Control boundary

At publication level, the control boundary has four roles:

1. estimate the expected behavioral effect of candidate memories;
2. choose among eligible candidates using task fit, policy constraints, and uncertainty;
3. compare observed behavior after recall with the predicted effect;
4. record divergences and support recovery actions when recall appears harmful or stale.

The key scientific claim is not that any one estimator is privileged. It is that a memory system becomes more testable when prediction and observation are placed on the same axis. Retrieval can then be evaluated not only by whether a memory was topically appropriate, but by whether the resulting behavior matched the intended intervention.

Lineage matters because memory systems accumulate state. Organizations contain superseded decisions, partial summaries, sensitive notes, model-generated reflections, and records that are true only under a historical context. Without recovery semantics, the memory layer can become a persistent source of confident drift.

Evidence status

This is an architecture-and-analysis preprint. It makes no empirical claim that Jacobian-causal control improves accuracy, safety, latency, or cost in deployed systems. Those are hypotheses.

The strongest version of the hypothesis is comparative: under matched models, memory stores, policies, and cost budgets, an influence-aware memory controller should reduce harmful stale-memory effects and improve post-error recovery relative to relevance-only retrieval, graph recall without influence observation, or policy filtering without post-intervention measurement.

The weaker but still useful claim is methodological: even when influence-aware control does not improve task outcomes, it may expose failure modes that relevance scores hide.

Assumptions

The public formulation assumes a fixed model or explicitly versioned model family during measurement, a comparable probe protocol, measurable behavioral dimensions, and memory interventions that are separable enough for local estimation to be useful. It also assumes that recovery primarily affects future system state. It cannot undo user decisions, external tool calls, or downstream social effects that have already occurred.

Disclosure boundary

The public materials intentionally avoid several categories of detail:

• production scoring functions, thresholds, and calibration procedures;
• private memory schemas, corpora, or evaluation traces;
• model-specific estimator choices;
• internal policy logic for eligibility, suppression, or recovery actions;
• deployment constraints that would allow the system to be reconstructed from the essay.

The preprint defines the research object and evaluation target. The operational details belong in protected implementation work, controlled evaluations, and attorney-reviewed patent materials.

How to test the claim

A fair evaluation should compare influence-controlled memory against simpler retrieval and graph-memory baselines under matched models, memory stores, policies, and cost budgets:

1. dense or hybrid top-$k$ retrieval;
2. graph retrieval or memory-tool recall without influence signatures;
3. relevance retrieval plus policy filtering;
4. influence reranking without observation;
5. influence-controlled memory with observation, recovery evidence, reliability, and lineage.

The strongest tasks are the ones where memory can help and hurt: compliance-sensitive recommendations, temporal supersession, project-memory recall, contradiction handling, tool-use planning, multi-session agent work, and stale or adversarial memory injection.

The primary metrics should include target-effect alignment, side-effect magnitude, answer accuracy, provenance completeness, stale-memory false-positive rate, recovery precision, post-recovery answer quality, latency, token cost, and human audit time.

The hypothesis should narrow if simpler retrieval or graph baselines match the controller at lower cost. It should fail if observed-effect measurement cannot distinguish useful memory interventions from harmful ones, or if recovery semantics do not improve diagnosis or post-error repair.

The proposal is also weakened if effect estimates are unstable across equivalent prompts, if evaluator choice dominates the measured signal, or if observed-effect metrics fail to predict downstream harm and recovery better than relevance scores.

Failure modes

Influence-controlled memory introduces state, and state creates path dependence.

• Stale influence: a memory's expected effect can drift after model, prompt, or policy changes.
• Hidden side effects: a memory can produce the intended answer while shifting tone, risk appetite, or tool behavior in unwanted ways.
• Interaction effects: two individually safe memories can combine into an unsafe frame.
• Signature leakage: influence signatures may reveal sensitive information if they are stored or shared carelessly.
• Incomplete recovery: generated text, tool calls, and user decisions may escape the model boundary before future system state is corrected.
• Target misuse: a controller optimized for influence can become manipulative unless targets are constrained by user intent, policy, and law.
• Measurement bias: the choice of probes, behavioral dimensions, and evaluators can encode policy assumptions rather than neutral scientific facts.
• Non-compositionality: local influence estimates may not compose when memories interact, prompts vary, or decoding stochasticity changes the effective intervention.

These risks are not incidental. They are the reason prediction, observation, and lineage must be explicit. A memory system that cannot say why a memory became influential is difficult to debug and difficult to govern.

Relation to adjacent work

This work sits near retrieval, agent memory, mechanistic interpretability, model editing, and inference-time control. It does not replace those fields. It asks for a missing layer between them: a memory controller that treats recall as an effectful operation.

Retrieval-augmented generation formalized the use of non-parametric memory to improve knowledge-intensive language generation, including provenance and updateability concerns (Lewis et al., 2020). ACT-R provides a cognitive-architecture precedent for treating memory availability as state- and goal-conditioned rather than simple lookup (Anderson et al., 2004). Influence functions and causal tracing show adjacent ways to ask how data or internal states affect model behavior, though they address different intervention sites and evidence regimes (Koh and Liang, 2017; Meng et al., 2022).

Retrieval asks which memories match the request. Influence control asks which memories should be allowed to change the model's behavior, how much confidence a system should place in that expected change, and what recovery path exists if the prediction is wrong.

For short-lived question answering, that may be unnecessary overhead. For long-running agents and AI-native organizations, it may become part of the reliability contract.

---

Publication

Marvin Danig, Causal Influence Control for Persistent Memory in Language Model Systems (2026). arXiv identifier and link: pending submission.

Evidence status. This architecture-and-analysis preprint reports no benchmark measurements. Proposed performance effects are hypotheses to be evaluated by the protocol above.

Materials availability. This essay introduces no experimental dataset or benchmark code. Public source and PDF artifacts are linked above.

Competing interests. The author is affiliated with Achiral Research. Certain aspects of the described architecture may be covered by patent applications associated with Achiral.