An AI memory proposal asks what should matter now—not just what matches

Why was a project canceled? A search might find the announcement and several related reports without recovering the sequence behind the decision. In a hypothetical example, losing a customer leads to a revised forecast, then a budget reduction, then cancellation. The records needed to explain that chain may not closely resemble the original question.

That distinction motivates activation-guided graph retrieval, or AGGR, proposed by Marvin Danig of Achiral Research. The November 2026 preprint describes an architecture and analyzes its costs. It reports no benchmark measurements and releases no implementation or experimental dataset. Its potential advantages remain hypotheses, not demonstrated improvements.

Matching a question versus informing an answer

Retrieval-augmented generation supplies a language model with information retrieved from an external store. Danig argues that selecting relevant records is only part of the problem for systems that retain information across many interactions. They must also decide which records should influence an answer now.

An old decision might match a question perfectly but have been superseded. A frequently retrieved summary might be less authoritative than the original observation. A record could be useful but inaccessible to the requester. AGGR treats these distinctions as part of memory availability rather than relying on similarity alone.

The proposed activation score combines query relevance with factors including successful past use, elapsed time, importance, active goals, source authority and penalties for staleness or contradiction. These contributions should remain inspectable. The score is an engineering ranking device, not a calibrated probability or evidence that the system reproduces human cognition. Its activation model draws inspiration from the ACT-R cognitive framework.

Following evidence as it appears

The architecture stores memory as a graph: records are nodes, and typed connections describe relationships such as temporal order, dependency, contradiction, provenance or causality. Request and goal cues identify starting points. Activation then helps allocate a limited search budget among possible paths.

A retrieved record can introduce new cues that were absent from the question. A cancellation record might lead to a budget review, which points toward a forecast and then a customer event. This recursive cue generation is intended to reach evidence outside an initially selected candidate pool.

Broadly connected cues receive reduced traversal weight. A term such as “roadmap” may connect to many records without helping distinguish the relevant ones. This degree-sensitive weighting aims to prevent generic connections from consuming the search budget, although excessive suppression could also hide useful routes. Tracking visited nodes prevents cycles from producing repeated expansion.

Authorization must happen before content-bearing scoring or expansion. Even connection counts and activation statistics must use only the requester’s authorized graph. Otherwise, restricted records could influence visible rankings and reveal information about their existence.

The output is not simply a ranked reading list. It is an evidence graph retaining selected records, typed relationships, sources, retrieval paths, score contributions, contradictory branches and unresolved gaps. The language model would receive this structure before generating its answer.

What the analysis establishes—and does not

The paper derives search-cost bounds under explicit limits on traversal depth, frontier size and neighbors examined. These limits constrain graph exploration; they do not make searching free. If a language model extracts new cues, its latency and token use are additional costs outside the graph-selection bound.

Nor does a structured evidence graph necessarily use fewer context tokens than retrieved passages. Node summaries, relationship labels and provenance all take space. The proposed benefit is better-supported relational evidence, which must be measured alongside its cost.

The architecture builds on existing activation, graph retrieval and persistent-memory approaches. Its claimed contribution is their ordered combination: state-dependent activation, degree-aware traversal, recursive cues, evidence reconstruction and reinforcement following qualified outcomes. Merely retrieving a record should not strengthen it; later confirmation, citation, successful action or consistency evidence should qualify that update.

A proposal with a clear test

The evaluation plan compares AGGR with lexical and dense retrieval, graph retrieval without activation, fixed-pool activation reranking and periodically consolidated notes. Tests would examine causal and temporal recovery, contradictions, provenance, stale evidence and answer accuracy, alongside latency, tokens and audit effort.

The central claim would be narrowed or falsified if a simpler graph reranker matched full traversal on causal and temporal recovery under equal model, evidence and cost budgets.

Graph quality remains a fundamental vulnerability. If an extracted relationship mistakes temporal sequence for causation, traversal can reconstruct a confident but incorrect explanation. Other risks include reinforcing popular records, preserving obsolete authority, treating model-generated reflections as observations, following an incorrect goal and producing graphs too large to audit.

The intended setting is therefore long-lived agents with changing goals, evidence and permissions—not necessarily simple questions over a small, stable collection. The preprint supplies a design and a way to challenge it. Whether that design earns its overhead remains unanswered.