Living Memory

1. Introduction

Retrieval-augmented generation (RAG) has become the default memory substrate for large-language-model systems, and its core abstraction has barely changed: embed documents, embed the query, return nearest neighbors. This abstraction is a warehouse, complete, indifferent, and static. Warehouses have three deficiencies as a substrate for a system that is supposed to learn:

1. No memory strength. A record retrieved ten thousand times and a record never retrieved once are identical to the index. There is no notion of consolidation, salience, or earned importance, properties that half a century of memory research identifies as central to intelligent recall.

1. No exploration. Nearest-neighbor retrieval is purely exploitative. The long tail of a knowledge base, often the majority of its mass, is structurally unreachable unless a query happens to land on it. Systems built this way converge on a small working set and starve their own periphery; novelty becomes an accident.

1. No schema formation. Most fundamentally: retrieval never changes the system. An insight surfaced today does not alter how tomorrow's information is encoded, organized, or perceived. The system accumulates but does not become.

Cognitive science offers well-tested accounts of all three dynamics, base-level activation in the ACT-R tradition (Anderson & Schooler, 1991), explore-exploit formalisms from the bandit literature (Auer et al., 2002), and schema-dependent consolidation from systems neuroscience (Tse et al., 2007; McClelland et al., 1995). The engineering contribution of this work is not any single mechanism but their composition into a coherent memory system for AI agents, together with the governance layer that composition demands, because a memory that evolves is a memory that can corrupt, and most of the design effort reported here concerns making evolution safe, bounded, and auditable.

We built this architecture for a specific and demanding application: a long-lived research intelligence whose corpus spans many domains (climate systems, institutional design, AI ethics, cooperative economics, consciousness research) and whose purpose is to find connections across them. Cross-domain insight is not a nice-to-have for this system; it is the product. That forcing function shaped every mechanism described below.

2. Related Work

Hierarchical retrieval. RAPTOR (Sarthi et al., 2024) demonstrated that recursive abstractive summarization trees outperform flat chunk retrieval. Our corpus organization anticipates this independently: knowledge exists simultaneously as atomic records, thematic consolidations, domain syntheses, and an executive summary, with retrieval able to answer at the appropriate altitude. We add what tree-RAG lacks: the hierarchy is dynamic, consolidation is proposed by the memory system itself from the evolving shape of its contents.

Cognitive architectures. ACT-R's declarative memory (Anderson & Schooler, 1991; Anderson et al., 2004) established base-level activation, a power-law-decaying, accumulating function of access history, as a model of human memory strength with unusual empirical reach. SOAR's chunking mechanism established learned structures that alter future processing. We import the form of these results into a modern embedding-retrieval stack; to our knowledge this composition is not present in deployed RAG systems.

Agentic memory systems. MemGPT (Packer et al., 2023) contributed paged memory tiers; generative agents (Park et al., 2023) contributed a retrieval function scoring recency × importance × relevance. These are single-decay heuristics without accumulation, saturation, exploration, or schema formation. They locate the right problem; we believe the mechanisms reported here go materially further toward solving it.

Multi-resolution embeddings. Matryoshka Representation Learning (Kusupati et al., 2022) produces embeddings whose prefixes are valid lower-dimensional embeddings, enabling coarse-to-fine retrieval funnels. We use MRL-class models both for scale (a cheap prefix scan over everything, full-dimension re-ranking over a shortlist) and, more distinctively, as one half of a dual-space representation described in §6.

Analogy and conceptual combination. Structure-mapping theory (Gentner, 1983) established that deep analogy maps relations, not attributes; conceptual blending (Fauconnier & Turner, 2002) described emergent structure from cross-domain integration. Our insight-detection layer is an operationalization of exactly this distinction: it is designed to find relational rhyme while explicitly discounting surface-vocabulary resemblance.

Exploration. Upper-confidence-bound methods (Auer et al., 2002) and diversity-aware re-ranking (Carbonell & Goldstein, 1998) are standard in recommendation and bandit settings but essentially absent from knowledge-base retrieval, where pure exploitation remains the norm. We argue this absence is a first-order defect in any memory system intended to support discovery, and we treat exploration as a budgeted, labeled component of every retrieval.

Minimum description length. Rissanen's MDL principle (1978) supplies our objective test for schema promotion: a learned schema is real to the extent that adopting it compresses the corpus.

3. Design Principles

Five commitments, stated up front because every subsequent mechanism is derived from them:

P1, The corpus is truth; everything else is derived. Source knowledge lives as human-readable documents under version control. Every index, embedding, graph, and score is disposable and re-derivable. This is not an implementation convenience; it is what makes evolution safe (§7.4).

P2, Memories are never deleted. Consolidation, decay, and refutation operate on interpretation and salience, never on existence. The unretrieved tail is a seed bank, not garbage.

P3, Relevance dominates; everything else tilts. All learned signals, strength, structure, priors, enter ranking as bounded adjustments over a relevance-ranked shortlist. No accumulation of secondary signals can invert a relevance judgment. (Our diagnostic work on a predecessor system found exactly this inversion, structural priors overwhelming content relevance until results were effectively query-independent, and we now treat boundedness as a design invariant rather than a tuning outcome.)

P4, The engine detects; the mind articulates; humans decide. Mechanical layers propose. Cognitive layers (LLM passes) articulate and attack. Promotion of anything into interpretive machinery requires explicit human consent, enforced in code rather than in policy documents.

P5, The history is the state. All learned state is stored as append-only event logs. Current state is a replay of history, which means every belief the system holds about its own memory is reconstructible and auditable at any past point.

4. Architecture Overview

The system comprises four layers:

1. Corpus, versioned human-readable documents organized in a consolidation hierarchy (atomic records → thematic clusters → domain syntheses → executive synthesis), with typed provenance links (parent, evidence sources, lateral connections) forming a navigable graph.

1. Derived indexes (disposable), a normalized catalog produced by a liberal-reader/strict-linter ingestion contract that accounts for 100% of corpus files with zero silent drops; content embeddings under an MRL-class model enabling a coarse-to-fine retrieval funnel; a lexical index; a link-graph centrality measure.

1. Learned state (append-only), the retrieval event log; the hypothesis ledger (§7.1); the learned-schema registry (§7.2); structural signatures (§6).

1. Retrieval and integration services, hybrid search (semantic funnel ∥ lexical, blended, then bounded re-ranking with strength and exploration terms); graph traversal along provenance; and a periodic offline integration cycle (§6.3) that detects candidate insights, updates the hypothesis ledger, and queues work for cognitive and human review.

We omit index parameters, blend weights, funnel dimensions, and profile definitions by policy; none are required to evaluate the architectural claims.

5. Memory Dynamics

5.1 Strength: memories earn their keep

Every retrieval event is logged. From this log the system computes a per-record activation in the ACT-R family: accumulating over the full access history, decaying by a power law, and saturating logarithmically. Three properties matter architecturally:

• Accumulation distinguishes chronic importance from momentary spikes, many accesses long ago and few accesses yesterday can weigh alike.
• Power-law decay means nothing ever fully vanishes; dormancy is a spectrum, not a state flip.
• Saturation means the ten-thousandth access adds almost nothing. Popularity self-limits by construction, the first of four defenses against the failure mode of §5.3.

We add a refinement not present in stock ACT-R: access diversity weighting. The system records the shape of each retrieving query and discounts strength earned from repetitions of the same question. A record retrieved by fifty differently-shaped questions is a load-bearing hub; a record retrieved fifty times by one question is a habit. Only hubs earn full strength. This single distinction separates the two phenomena that raw frequency conflates, and empirically it is the difference between an activation signal that identifies understanding and one that identifies ruts.

Activation tilts ranking within the P3 bound and never gates storage (P2).

5.2 Exploration: the seed bank gets its day in court

A fixed, mode-dependent fraction of every result set is reserved for wildcards: records that pass relevance screening but lie in the dormant band of the activation distribution. Wildcard selection is not random. Candidates carry an uncertainty bonus in the upper-confidence-bound family, proportional to how under-observed the record is, decaying as it accrues examinations. The system is not injecting noise; it is systematically paying down uncertainty about its own long tail, with the convergence behavior that family of methods guarantees.

Two disciplines make this workable in practice:

• Wildcards are labeled. The consuming agent always knows which results are canon and which are provocation. Unlabeled mixing teaches the consumer to distrust retrieval; labeled wildcards become an invitation.
• Exploration temperature follows mode. Precision-critical retrieval runs at zero temperature; synthesis and discovery modes run warm; the offline integration cycle runs hot, deliberately pairing high-activation records with dormant ones, the architectural analog of REM-sleep replay, which biological systems disproportionately spend on non-habitual trajectories.

A diversity constraint at result selection (in the maximal-marginal-relevance family) additionally prevents any tight cluster from monopolizing a result set.

5.3 The gravity-well problem

Use-weighted retrieval has an obvious pathology: rich-get-richer feedback. Frequently retrieved records rank higher, get retrieved more, rank higher still, and the system spirals into a small attractor set (the psychological analog is rumination; the sociological one is ideology). We consider this the central failure mode of learned retrieval, and the architecture defends it from four directions simultaneously: saturation flattens wells from above (§5.1); diversity weighting starves them from the side (§5.1); uncertainty bonuses pump from below (§5.2); and the hot integration cycle stirs the reservoir on a schedule (§6.3). Our design position is that no single mechanism suffices and that the four must be present together; each covers a failure mode the others miss.

6. Emergent Insight: Two Spaces and a Differential

6.1 Why one embedding space cannot find deep analogy

The cross-domain insights worth having are structural: threshold-triggered cascade dynamics appearing in both reef ecology and interbank contagion; annealing schedules rhyming with creative incubation; immune tolerance rhyming with content moderation. Text embeddings, trained on co-occurrence, place such pairs far apart, because their vocabularies are disjoint. Conversely, what clusters closely in content space is topical overlap: the same subject discussed in two settings. A detector built on content similarity therefore surfaces shallow resonance and is structurally blind to deep resonance. This is not a tuning problem; it is a representation problem.

6.2 The dual-space differential

Our solution is a second representation. Each record is distilled, by a cognitive (LLM) pass operating under a protocol we do not publish, into a domain-neutral structural signature: the record's variables, feedback signs, thresholds, and dynamics, with all domain vocabulary removed. Signatures are embedded in their own space. Candidate insight is then scored by a differential: structural similarity discounted by content similarity. Pairs that are close in both spaces are topical neighbors (unremarkable); pairs close in structure and far in content are built the same while talking about different worlds, which is, we contend, the operational definition of deep analogy, and a direct operationalization of structure-mapping theory's relations-not-attributes criterion.

The distillation protocol is the most leverage-dense component of the system, and its quality ceiling is an open empirical question we are actively characterizing. The abstraction ladder extends downward, content → phenomenon → structure → formal skeleton, and at the deepest rung, signatures approach shared mathematics, where rhyme is strongest and the same memory substrate becomes a working tool for scientific and design problems.

6.3 The integration cycle

On a schedule (the system's "sleep"), an offline process: (i) detects candidate clusters in both spaces; (ii) scores cross-domain candidates by the differential; (iii) reconciles detections against the hypothesis ledger (§7.1), recording trajectories; (iv) deliberately probes high-activation × dormant pairings across domain boundaries; (v) refreshes graph centrality; and (vi) emits a review queue for cognitive and human attention. Detection thresholds are calibrated against a null model (domain-label shuffling) so that surfaced candidates must beat the spurious-cluster distribution, necessary because false-discovery count grows superlinearly with corpus size while real structure grows roughly linearly, and an insight engine that drowns its operators in coincidence has failed at scale even if every mechanism works.

7. Insight Promotion and Governance

A detected insight is a hypothesis. The distance from hypothesis to interpretive machinery, a learned schema that changes how future information is encoded, is where an evolving memory either becomes a learning system or a self-radicalizing one. This section is the core of the work.

7.1 The hypothesis ledger: refutation demotes, never deletes

Promotion false positives are expensive (a wrong schema distorts everything encoded under it); rejection false negatives are cheap if and only if rejection is non-destructive. We therefore make rejection non-destructive and let the gate be strict:

• Every candidate ever detected is retained permanently with full history.
• Adversarial review, independent cognitive passes explicitly prompted to break the candidate, through heterogeneous lenses (structural soundness, evidence independence, vocabulary-coincidence), issues graded verdicts. A refuted candidate becomes dormant, not dead.
• Detection cycles match new candidates against the ledger by membership overlap. When new evidence independently arrives in a dormant candidate's neighborhood, it resurfaces carrying its prior refutations as questions it must now answer.
• A candidate that returns across cycles with rising scores and independently-arriving members exhibits precisely the signature of a real connection blooming slowly, and that trajectory is admitted as promotion evidence stronger than any single-cycle score.

The ledger yields an empirical gauge of the system's own skepticism: the recovery rate, the fraction of eventually-promoted schemas that were previously refuted. A high recovery rate means the gate is too aggressive; the threshold is corrected from measurement rather than intuition. The system tunes its own doubt.

7.2 The schema lifecycle: promotion by prediction, not detection

Learned schemas advance through a typed state machine: candidate → trial → established → deprecated. Two transitions carry the architectural weight:

• Candidate → trial is evidence-gated: differential score, survival of adversarial review, ledger trajectory.
• Trial → established requires explicit human consent, enforced in code. The transition is mechanically impossible without a named approver. This is P4 made structural: the one element of the promotion loop that cannot be captured by the loop is a decision-maker outside it.

Trial schemas must earn establishment: measurable retrieval lift from the event log; correct anticipation of structure in records ingested after proposal (a schema that only postdicts is numerology); and compression, under the MDL criterion, a true schema reduces the description length of the corpus it claims to organize.

Two further commitments: learned schemas are deposable (deprecation is a first-class transition, an evolving system must be able to recover from an early wrong turn), in explicit contrast to the small constitutional layer (the system's ethics and purpose), which is human-owned, append-only, and outside the learned lifecycle entirely. And influence is recorded: every artifact encoded while a schema is active carries that schema's identity and version, so causal diffusion through the corpus remains auditable indefinitely.

7.3 What planting a schema changes

An established schema participates in interpretation four ways: as taxonomy (a tag future records can carry), as an ingestion annotator (new records are scored against schema signatures, flagging "this new document rhymes with schema X"), as a retrieval lens (search conditioned through a schema's structure), and as a consolidation prior (new clusters are tested for assimilation into existing schemas before accommodation creates new ones, the Piagetian distinction, made operational).

7.4 Checkpointed worldviews

Because content representations anchor to an immutable corpus (P1/P2) and schemas live entirely in derived layers, the system's whole interpretive state can be re-derived under any historical schema set and diffed. Biological memory cannot do this: schema change in brains silently rewrites the past (confabulation), with no audit trail. Here, worldview is a versioned artifact. For applications where trust in machine-formed beliefs is the binding constraint, which we believe includes every civilizationally consequential application, we regard this property as the decisive argument for the architecture, above any retrieval-quality metric.

8. Failure-Mode Analysis

We summarize the five failure modes that dominated design, and their standing mitigations:

Failure mode	Mechanism of harm	Mitigation
Gravity wells	use-weighting feedback → attractor working set	saturation + diversity weighting + uncertainty bonuses + hot cycle (§5.3)
Self-confirming schemas	promotion loop rewards its own outputs	prediction/compression gates, deposability, human consent outside the loop (§7.2)
False-discovery scaling	coincidence grows superlinearly with corpus	null-model thresholds, rate-limited surfacing (§6.3)
Provenance dilution	schema influence becomes causally untraceable	influence stamping, append-only event state (§7.2, P5)
Shallow-insight capture	content-space detectors reward vocabulary overlap	dual-space differential (§6.2)

We regard the joint presence of these mitigations as the contribution; each is individually insufficient, and several are individually known.

9. Implementation Status and Observations

All mechanisms described are implemented and exercised end-to-end against a living multi-domain research corpus (~750 documents, ~7,500 retrieval units) that had previously been operated under a metadata-only retrieval system. Verified behaviors include: 100%-accounted ingestion with per-field drift reporting on a corpus that had drifted substantially from its own contract; activation accumulation with diversity weighting from live retrieval logs; labeled wildcard selection under mode-dependent temperatures; the full hypothesis-ledger loop including refutation → dormancy → evidence-driven resurfacing with trajectory → consent-gated promotion, with the recovery-rate gauge reporting correctly; and dual-space detection correctly separating a structurally-rhyming cross-domain triple (differential strongly positive) from topical neighbors (differential ≈ 0) in controlled tests. Full-scale semantic evaluation under production embedding models, threshold calibration against null models, and the first at-scale distillation pass are in progress; we report them in a subsequent edition rather than projecting results we do not yet have.

10. Disclosure Statement

This preprint describes architecture and rationale at the level required to evaluate the claims and to situate the work in the literature. We withhold: scoring compositions and all numeric parameters; profile and temperature definitions; funnel geometry; ledger-matching and calibration procedures; the ingestion contract's normalization rules; and the structural-signature distillation protocol in its entirety. These withholdings are commercial, not epistemic: none affect the validity of the architectural argument, and each is independently replaceable by a determined implementer, what they jointly protect is the calibration that makes the composition work in practice.

11. Limitations

The distillation quality ceiling is unproven at scale; signatures are produced by LLM passes whose consistency across domains is an open question. MDL-based compression gating is specified but not yet operationalized as a routine metric. The exploration stack's long-run behavior under adversarial or heavily skewed query distributions has not been characterized. Human consent as the promotion backstop assumes an attentive operator; the architecture reduces but does not eliminate governance burden. And the deepest claim, that machine-detected structural rhyme, governed as described, yields genuinely novel insight at a rate that justifies the machinery, is exactly the claim our ongoing deployment exists to test.

12. Conclusion

A warehouse remembers everything and learns nothing. A brain learns continuously and can neither audit nor version its own beliefs. The architecture described here occupies deliberately chosen middle ground: memory that strengthens with use but saturates; that explores its own dormant tail on principle rather than by accident; that detects when distant regions of knowledge are built the same; and that changes its own perception only through a gate whose final position is held by a human being, with every step recorded, every influence stamped, and every worldview reconstructible. We believe this is the correct shape for machine memory in consequential domains, and we offer the composition, not any single mechanism, as the contribution.

---

References

• Anderson, J. R., & Schooler, L. J. (1991). Reflections of the environment in memory. Psychological Science, 2(6).
• Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Qin, Y. (2004). An integrated theory of the mind. Psychological Review, 111(4).
• Auer, P., Cesa-Bianchi, N., & Fischer, P. (2002). Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47.
• Carbonell, J., & Goldstein, J. (1998). The use of MMR, diversity-based reranking for reordering documents and producing summaries. SIGIR '98.
• Fauconnier, G., & Turner, M. (2002). The Way We Think: Conceptual Blending and the Mind's Hidden Complexities. Basic Books.
• Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2).
• Kusupati, A., et al. (2022). Matryoshka representation learning. NeurIPS 2022.
• McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. Psychological Review, 102(3).
• Packer, C., et al. (2023). MemGPT: Towards LLMs as operating systems. arXiv:2310.08560.
• Park, J. S., et al. (2023). Generative agents: Interactive simulacra of human behavior. UIST 2023.
• Rissanen, J. (1978). Modeling by shortest data description. Automatica, 14(5).
• Sarthi, P., et al. (2024). RAPTOR: Recursive abstractive processing for tree-organized retrieval. ICLR 2024.
• Tse, D., et al. (2007). Schemas and memory consolidation. Science, 316(5821).