Activation Is Not Querying

The distinction between querying and activation matters because a larger context window, a vector index, and a knowledge graph are all useful substrates, but none of them specifies a theory of memory availability. They answer where information resides and how a system can search it. They do not, by themselves, answer which prior experiences should exert influence on the system now. Querying exhibits explicit search behavior with total recall, while activation exhibits implicit influence based on prior engagement or forgetfulness.

The retrieval question is underspecified

Retrieval-augmented generation typically begins with a query $q$, a collection of stored objects $M = \{m_1, \ldots, m_n\}$, and a relevance function:

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

The system returns the top-$k$ objects under $s_q$, possibly after lexical fusion, metadata filtering, or learned reranking. Graph RAG broadens the retrievable unit from a passage to a node, neighborhood, path, community, or summary. These are important advances. They improve candidate generation and expose relationships that flat passage retrieval suppresses.

But the mathematical object remains mostly query-conditioned relevance. A record can score highly whether it was confirmed yesterday or superseded last year; whether it represents a direct observation or a model-generated reflection; whether it has repeatedly supported successful action or has merely been retrieved many times; and whether it belongs to the current goal state or happens to share vocabulary with it. Each defect can be repaired with another filter or feature, but collectively they point to a missing variable: availability as a function of the system's history and present cognitive state.

We call that variable activation.

Activation versus querying

Querying and activation are complementary operations over different conditionals.

Querying estimates something like:

$$P(m_i \text{ is relevant} \mid q).$$

Activation should estimate a richer quantity:

$$P(m_i \text{ should influence inference now} \mid q, g, H_t, G, \pi, \sigma_i),$$

where $g$ is the active goal or workflow state, $H_t$ is interaction and access history, $G$ is memory-graph structure, $\pi$ is the permission policy, and $\sigma_i$ is the epistemic state of the memory: source, confidence, verification, contradiction, lineage, and decay status.

This is not a terminological distinction. It changes the retrieval operator.

Dimension	Query-conditioned retrieval	Activation-conditioned memory
Primary question	What resembles or answers this request?	What prior experience should be cognitively available now?
State	Commonly request-local	History-, goal-, and workflow-dependent
Unit of selection	Passage, document, row, node	Entity, episode, procedure, edge, path, or evidence subgraph
Temporal behavior	Usually static between indexing updates	Strengthens, decays, and changes with use and outcome
Structural behavior	Ranks candidates independently or by neighborhood	Allocates traversal budget across connected evidence
Epistemic behavior	Often delegated to downstream generation	Carries provenance, contradiction, and verification into selection
Output	Ranked context items	A bounded, auditable evidence graph

A vector score should therefore be understood as one term in a memory policy, not as the policy itself. Querying nominates candidates. Activation governs which candidates and paths are allowed to become influential.

A stateful activation operator

The filed architecture represents organizational experience as a hierarchical graph containing goals, domains, episodes, events, chunks, facts, entities, procedures, decisions, source documents, and agent actions. Edges may encode temporal order, causality, containment, contradiction, update, provenance, derivation, similarity, assignment, or dependency.

For memory entity $m_i$, a minimal activation model can be written as:

$$B_i(t) = \log(n_i + 1) - d_i \log(\Delta t_i + 1) + \beta S_i,$$

where $n_i$ is a qualified access or successful-use count, $d_i$ is a memory-specific decay rate, $\Delta t_i$ is time since a qualifying event, and $S_i$ is salience. This is an engineering approximation inspired by base-level activation in ACT-R, not a claim that an organizational memory system is a faithful cognitive model.

At retrieval time, the base term can be combined with query relevance, spreading activation, utility, epistemic status, and policy:

$$A_i = \eta s_q(m_i) + B_i(t) + \sum_j W_j S_{ji} + U_i + V_i - C_i + \epsilon,$$

where $W_j S_{ji}$ represents activation spread from cue $j$, $U_i$ estimates task utility, $V_i$ represents verification or source authority, and $C_i$ penalizes contradiction, staleness, redundancy, or policy risk. Hard authorization constraints must be applied before content or relevance can leak into the response; permission is not merely another soft feature.

This decomposition creates useful research degrees of freedom. The terms may be hand-calibrated, learned, or combined by a constrained policy. More importantly, each term can be ablated and audited. When a memory enters context, the system can explain whether it arrived because of semantic match, recency, repeated successful use, a causal edge, an active goal, or source authority.

Activation changes graph traversal

The most consequential part of the architecture is not activation reranking over a fixed candidate list. It is using activation to control where retrieval goes next.

Given initial cues extracted from a query, goal, or agent state, the system activates seed nodes and constructs a bounded frontier. It then expands candidate paths using a fanning score such as:

$$F_{ij} = w_j \cdot a_{ji} \cdot \phi(\deg(j)) \cdot \psi(A_i),$$

where $w_j$ is cue importance, $a_{ji}$ is association strength, $\phi$ attenuates high-fan cues, and $\psi(A_i)$ maps node activation into traversal priority. One possible attenuation is:

$$\phi(\deg(j)) = \frac{1}{1 + \alpha(\deg(j)-1)^2}.$$

The fan term addresses a familiar enterprise failure mode. Nodes such as “customer,” “roadmap,” “Q3,” or “incident” may connect to thousands of memories. Their raw connectivity makes them easy to retrieve and nearly useless for discrimination. Fan attenuation preserves them as cues while reducing their power to consume the retrieval frontier.

Paths can then be scored by accumulated evidence value less traversal cost and redundancy:

$$\operatorname{PathScore}(p) = \sum_{(j,i) \in p} F_{ij} - \lambda \operatorname{Cost}(p) - \gamma \operatorname{Redundancy}(p).$$

Retrieved nodes recursively emit new cues. A budget review can activate a revenue forecast; the forecast can activate an account loss; the account loss can activate a renewal-risk decision. This recursive cue generation is how the system can recover a causal chain whose decisive evidence does not resemble the original wording.

The output is not an unordered top-$k$ list. It is an evidence graph that preserves the retrieved nodes, typed edges, path history, provenance, activation contributions, confidence, contradictory branches, and unresolved gaps before generation begins.

A concrete failure of query-only recall

Consider the question: Why was Project X canceled?

A strong embedding model may retrieve the cancellation announcement, a meeting transcript containing “Project X,” and a status report semantically close to “canceled.” Those passages could support a fluent answer while omitting the actual cause.

In the activation-guided formulation, “Project X” and “canceled” seed the graph, but they do not determine the answer. The cancellation decision activates its decision record and temporal predecessors. The decision record fans into the budget review. The budget review activates the revised revenue forecast. The forecast activates the loss of a specific account. Evidence reconstruction yields a typed chain:

$$\text{account loss} \rightarrow \text{forecast revision} \rightarrow \text{budget reduction} \rightarrow \text{cancellation decision}.$$

The language model receives that chain together with source provenance and any competing explanation. Generation is downstream of retrieval-time reconstruction, rather than being asked to invent causal structure from adjacent fragments.

This example exposes the key research claim: for questions requiring temporal, causal, procedural, or decision lineage, the relevant unit is often a path under a goal-conditioned memory policy, not a passage under a similarity metric.

What follows analytically

Three design-level conclusions follow from the formulation.

1. Activation can bound a graph frontier

Naive breadth expansion with average branching factor $b$ and depth $h$ has worst-case growth $O(b^h)$. If activation retains a frontier of at most $m$ candidates per depth, expands at most $b$ neighbors per retained candidate, and maintains the frontier with bounded selection, a more honest implementation estimate is:

$$O(hmb + hmb\log m),$$

or $O(hmb)$ with linear-time bounded selection. The precise constant depends on index layout, batching, and whether scores are cached. Activation does not abolish graph-search complexity; it supplies a principled pruning policy. Earlier shorthand estimates that omitted the multiplicative frontier width understated the work.

2. Evidence density matters more than token count

For $k$ independent chunks averaging $c$ tokens, flat context cost is approximately $kc$. An evidence graph with $r$ retained node summaries of length $s$, $|E|$ serialized edges of cost $e$, and provenance cost $p$ requires approximately:

$$T_{graph} \approx rs + |E|e + rp.$$

This representation is not guaranteed to be shorter. Its proposed advantage is higher evidence density: more causal, temporal, and epistemic structure per token, with less duplicated passage text. A fair evaluation must report both token cost and support coverage.

3. Reinforcement introduces a feedback-control problem

For $r$ retrieved memories, metadata reinforcement can be implemented with $O(r)$ asynchronous updates. Computationally, that is modest. Statistically, it is dangerous. If retrieval itself increments future availability, an early false positive can become self-reinforcing.

Reinforcement should therefore depend on qualified outcomes - citation, user confirmation, successful action, or later consistency - rather than retrieval alone. Negative evidence and supersession must reduce authority without destroying lineage. In this architecture, forgetting is not deletion; it is controlled loss of retrieval authority.

Testable hypotheses

The architecture should earn its complexity empirically. We propose comparing five systems under matched generation models and context budgets:

1. lexical retrieval;
2. dense or hybrid top-$k$ retrieval;
3. graph retrieval without activation;
4. activation reranking over a fixed candidate pool;
5. activation-guided traversal with recursive cue expansion and evidence reconstruction.

The strongest evaluation tasks are those where flat similarity is plausibly insufficient: organizational “why” questions, decision-history reconstruction, incident postmortems, temporal supersession, customer-commitment recall, procedural next-action selection, and contradiction detection.

Primary metrics should include answer accuracy, evidence-node recall, causal-edge F1, temporal-order accuracy, contradiction recall, provenance completeness, stale-memory false-positive rate, and calibration of answer confidence against evidence support. Systems should also report retrieval latency, graph operations, serialized context tokens, and human audit time.

The critical ablations remove, one at a time, base activation, goal state, fan attenuation, recursive cue generation, epistemic weighting, evidence-graph serialization, reinforcement, and pre-retrieval authorization. The claim is falsified or narrowed if a simpler graph reranker matches full traversal on causal and temporal recovery at equal cost.

Failure modes and boundary conditions

Activation is not intrinsically safer or more accurate than querying. It creates state, and state creates path dependence.

• Popularity lock-in: frequently retrieved memories can crowd out rare but decisive evidence.
• Stale authority: recency and repetition can preserve an obsolete decision unless supersession edges carry sufficient weight.
• Reflection laundering: model-generated summaries can acquire the authority of direct observations if lineage is lost.
• Fan overcorrection: strong attenuation can suppress broad concepts that are genuinely diagnostic for a task.
• Goal capture: an incorrect active goal can steer traversal away from relevant evidence.
• Permission leakage: scoring an inaccessible node before authorization may reveal that it exists, even if its content is withheld.
• Audit overload: a large evidence graph may be more inspectable in principle but less usable in practice.

These are reasons to expose activation contributions, provenance, and path history as first-class outputs. A memory system that cannot explain why an experience became available is difficult to calibrate and unsafe to reinforce.

Relation to adjacent work

This proposal does not claim to invent activation, graph retrieval, hierarchical memory, reflection, or retrieval-augmented generation. ACT-R provides a mature account of base-level and spreading activation. RAG established retrieval from non-parametric memory for generation. GraphRAG demonstrated graph-based, query-focused summarization. RAPTOR developed recursive tree-organized retrieval. Generative Agents combined memory streams, reflection, and planning. MemGPT framed LLM memory management through an operating-system analogy, and Zep developed temporal knowledge-graph memory for agents.

The proposed contribution is the ordered runtime combination: goal- and history-conditioned activation; fan-aware traversal of a hierarchical organizational memory graph; recursive cue generation; evidence-graph reconstruction before language-model inference; and auditable reinforcement after use. The patent context is relevant because it fixes the claimed engineering boundary. The research question remains open because the value of that combination must be established through controlled comparison.

The argument in one sentence

Querying asks which stored objects match the request; activation determines which prior experiences and evidence paths should be allowed to shape the system's present behavior.

For short-lived question answering, that distinction may add little. For agents and organizations expected to accumulate experience, revise beliefs, preserve decision lineage, and operate under policy over months or years, it may be the difference between a searchable archive and a memory system.

Publication

Achiral, Activation-Guided Graph Retrieval for Cognitive Memory Reconstruction in Language Model Systems (2026). arXiv identifier and link: pending submission. This page is the canonical companion essay and will be updated when arXiv assigns the record.

The work is associated with U.S. provisional patent application 64/084,217, Hierarchical Activation-Guided Graph Retrieval for Cognitive Memory Reconstruction in Large Language Model Systems, filed June 6, 2026.

Copyright 2026 Achiral. Patent rights reserved. No patent license is granted by publication of this article.