First documented 3D temporal KEG and selective-resolution memory model.
Memory is not a database.It is changing access to preserved meaning.
MMRY is an adaptive memory architecture that binds multimodal information, context and affect across time, preserving provenance while changing the resolution at which information is retained and recalled.
Bound across time without collapsing their unique properties.
Pattern-aware forecasts and provenance-bound discovery of unseen connections.
Preserve what information is.
Record what it means.
The information object remains unchanged and provenance-bound. A person’s emotional relationship to it is stored as a separate, evolving first-class object. Affect changes importance, consolidation, retrieval and decay—not informational truth.
MMRY is biologically inspired, not a simulation of the brain. It serves human, machine, organisational, media and agent memory without requiring every stored event to have been “lived.”
From evidence to an inspectable adaptive memory.
Click a layer to inspect its role.
Memory orchestrator
Segments episodes, resolves entities, preserves provenance and writes linked representations with epistemic status.
Evidence before interpretation
Information remains distinguishable from extraction, interpretation, prediction and reconstruction.
Bound, not flattened
Information, context and affect share an episode identity while preserving their unique structures.
Accessibility before erasure
Decay first changes surface access, resolution and retrieval probability; the compressed core may persist.
Authority is separated
Preventing decay never automatically grants authority to retrieve, disclose, alter or reinterpret.
The surface moves.
The compressed core holds.
What fades from immediate access may remain retained and reconstructable below the surface.
Temporal surface
Current, recent, frequent and contextually relevant information. High-resolution and immediately available.
VIVIDFrequent memory
Repeatedly accessed information, pre-indexed relationships and current affective context.
CLEARConsolidated memory
Graph relationships, semantic representations, GIST and affective imprints selectively assembled.
FADEDCompressed core
Highly folded information, provenance, associations and reconstruction instructions retained below ordinary recall.
HAZY → FORGOTTENDecay reduces the cost and likelihood of access before it reduces the existence of information. FORGOTTEN means absent from ordinary automatic recall—not necessarily erased. Explicit deletion remains a separate governed operation.
Some memories cannot decay automatically.
User-locked, safety-critical, policy-protected or administrator-specified memories preserve their required meaning, provenance and retrievability.
UserExplicitly protected or released by the person.
System policyProtected by defined safety or operational rules.
AdministratorAssigned by an authorised role with reason, scope and review.
THRPYCritical boundaries, crisis context, protective relationships and confirmed breakthroughs.
Authority to prevent decay is not authority to retrieve, disclose, alter or reinterpret the memory.
Architecture iterations, documented over time.
Project provenance from Timothy McGuckin’s MMRY record—not peer-reviewed validation. Implementation claims are shown conservatively where independent verification is incomplete.
Temporal Knowledge–Experience Graph
A responsive 3D graph of interrelated memory events; multimodal capture; selective high-resolution retention; reconstruction of faded context.
VIVID · CLEAR · FADED · HAZY · FORGOTTENEdge multimodal target
A Raspberry Pi, AI HAT and Hailo-accelerated target joined dual-camera capture, time-synchronised video/data, overlays, selective retrieval and secure transmission.
EDGE · VISION · TIME-SYNCChronology as the spine
MMRY was specified as a secure package of media, documents, video, metadata and knowledge graphs, playable as film or interactive timeline; time and sequence became its signature.
CAPSULE · 3D KEG · TIMELINEThe capsule is also its blueprint
The timeline capsule became both stored content and the instructions needed to index, selectively retrieve and regenerate it across semantic, temporal and emotional layers.
CONTENT + RECONSTRUCTION CONTRACTExact retrieval + personal data exchange
Requirements added lossless deterministic access to exact text and numbers in milliseconds, hot/warm/cold tiers, arbitrary personal data and consent-scoped PDX integration.
RANDOM ACCESS · PDX · GLYPHWAVETime-synchronised multimodal capsule
Architecture work joined video, audio, detections, event boundaries, graph, context and emotion inside a self-contained memory capsule with graph + vector retrieval.
MMRY → KEG → VECTOR RETRIEVALA wider cognitive system
MEMO Vision captures context; ReMindMap expresses the live thought graph; MMRY becomes compressed long-term contextual memory; THRPY supplies reflective interaction.
CAPTURE → THOUGHT → MEMORY → REFLECTIONAffect, protection and intelligence
SensAffect formalised Emotion Atom → Affective Imprint → Affective Presence. MMRY separated information from evolving affect, added protected memories, prediction and provenance-bound epiphanies.
INFORMATION × CONTEXT × AFFECTObservation → constraint → engineering response.
Memory traces change across encoding, consolidation, retrieval and forgetting.
Recent reviews describe engrams as dynamic rather than static records.[1][6]
Resolution states, not immutable rows
VIVID through FORGOTTEN models a governable information lifecycle—an analogy, not a cell-state map.
Engram sparsity is associated with precision; selective traces emerge over consolidation.
Animal studies support selective, distributed representation.[2][3]
Selective retention + linked representations
Preserve high-resolution evidence where it matters while distributing semantic, temporal and relational indices.
Sparse, event-specific hippocampal patterns distinguished individual food-caching events.
A 2024 Cell study described rapidly changing episode patterns.[4]
Episode identity + multimodal binding
Each capsule receives durable identity without requiring every modality to share one representation.
Forgetting can be an active, adaptive process supporting flexibility.
A 2024 review surveys molecular, cellular and network mechanisms; much evidence is animal-based.[5]
Controlled forgetting with policy
Decay, compression, masking and deletion are separate. User agency and legal deletion override inferred salience.
Emotion became a first-class data object.
Digitised affect is bound to information without becoming the information or rewriting its truth.
Atom → imprint → multimetric cloud
Dots orbit a 3D semantic sphere. Hover for labels. Move the VAD sliders to rebuild the cloud.
Illustrative, not diagnostic. SensAffect stores modality, uncertainty, time and provenance alongside every estimate. Adjust sliders — the cluster rebuilds below the atom.
Information object
What was captured, observed, asserted or generated. Evidence and provenance remain stable.
Affective object
How this information relates emotionally to a particular user, system or perspective across time.
Memory relationship
Contextual significance that influences importance, consolidation, resolution, retrieval and decay.
Emotion Atom
One bounded VAD estimate plus candidates, intensity, modality, confidence, perspective and provenance.
Affective Imprint
A complete Emotion Cluster Cloud for a bounded window, including conflicting multimodal atoms.
Affective Presence
Imprints evolving through recency, recurrence, salience, decay and reactivation without rewriting evidence.
MMRY’s emotional-indexing requirement became a reusable affective architecture for detection, interpretation, generation and governance. SensAffect digitises the relationship; MMRY binds it to information and decides how that memory is handled.
A useful coordinate system—with important limits.
Why VAD?
Valence represents pleasantness, arousal activation, and dominance perceived control. PAD/VAD originates in Mehrabian and Russell’s 1974 work; its interpretation remains debated.[7]
Continuous axes preserve gradients categorical labels collapse. Anger and fear can share negative valence and high arousal while differing in dominance.
Why multimodal?
Text, voice, facial behaviour and physiology expose different signals. MERSA combines speech, text, physiology and self-report and predicts VAD.[8]
A 2014–2024 systematic review finds recurring dataset, fusion, generalisation and real-world evaluation challenges.[9]
What it refuses to claim
Detection estimates expression and context; it does not directly observe private emotional state. Scores are probabilistic, culturally contingent and modality-dependent.
Mixed-emotion research reinforces retaining distributions and disagreement rather than forcing one label.[10]
Memory becomes prospective.
The same graph, vectors, semantic memory, time and affect that support recall can discover patterns without turning inference into fact.
Supporting memories, graph paths, similarity evidence, alternatives and confidence travel with every output.
Prediction
What may happen, matter or be needed next?
Epiphany
What previously separate memories become meaningful when connected?
ObservationCaptured directly
AssertionStated by a person or system
InterpretationMeaning derived from evidence
PatternRecurring relationship detected
PredictionProposed future state
EpiphanyProposed unseen connection
OutcomeWhat subsequently happened
MMRY remembers what happened, preserves what it meant, discovers what connects, and predicts what may matter next.
A confirmed epiphany may be protected from ordinary decay, but remains contestable, revisable, supersedable and deletable. A prediction is linked to its eventual outcome rather than silently rewritten as fact.
Claims you can follow back to source.
From research architecture to live applications.
Status noted 19 August 2026. Product-listed URLs are creator provenance, not independent performance validation.
THRPY
Voice-first emotional AI and reflection layer consuming affective context and protected memory.
thrpy.chat ↗PROJECT-LISTED URLMMRY Chat
Conversational interface for the adaptive memory architecture.
mmry.chat ↗PROJECT-LISTED URLPunctilio One
Adjacent governance system reusing affective and provenance patterns—not a redefinition of MMRY.
punctilio.one ↗Established. Inferred. Proposed.
EstablishedPeer-reviewed findings about memory dynamics and affect representation, linked to papers and reviews.
Engineering inferenceGraph distribution, adaptive fidelity, episode binding and controlled forgetting inspired by those findings.
Proposed by MMRYThe KEG, five resolution states, compressed core/surface model, affect binding, protected memories, prediction and epiphany objects.
Not claimedConsciousness, perfect recall, diagnosis, direct access to private emotion, or equivalence to a biological brain.
Adaptive AI memory —what MMRY proposes.
Long-term, multimodal memory for agents and applications: graph + vector recall, governed resolution, emotional context, and provenance.
- What is AI memory in the MMRY architecture?
- MMRY defines AI memory as adaptive, provenance-bound access to preserved meaning—not a static database. Multimodal evidence, context, and affect are bound across time with governed resolution from VIVID to FORGOTTEN.
- How is MMRY different from RAG or vector-only memory?
- MMRY combines a Knowledge–Experience Graph, vector retrieval, semantic memory, temporal trajectories, and affective imprints. Resolution states, protected memories, and epistemic status separate observation from inference.
- Does MMRY support multimodal AI memory?
- Yes. Voice, text, images, documents, activity, and sensor context enter as timestamped evidence inside memory capsules with graph and vector indices for selective recall.
- What is SensAffect in an AI memory stack?
- SensAffect is the affective layer: Emotion Atom, Affective Imprint, and Affective Presence. It stores how information relates emotionally without rewriting informational truth.
- Is MMRY a product or research architecture?
- This site documents a research and engineering architecture with lineage since December 2024, technical specifications, peer-linked evidence, and reference implementations. Live surfaces include mmry.chat and thrpy.chat.