1. Introduction
The dominant paradigm in AI assistants treats the system as a stateless function. A user submits a prompt. The system generates a response. Between requests, the system has no state, no experience, and no passage of time. It does not know whether it has spoken to this user before.
This is not a criticism. It is a design choice. But it raises a question: what happens when that constraint is relaxed? What does an agent look like when it maintains a persistent memory that evolves over time, decays with disuse, and strengthens with recall?
This paper describes the memory architecture of Potato, a local AI agent built to test that question. The memory system is not a database bolted onto a chatbot. It is the mechanism that makes the agent a continuous entity. The architecture has five interacting components: a sliding window decay model, a reconsolidation mechanism, a nightly sleep decay system, a geometric stickiness model derived from a three-dimensional affective state space, and a true forgetting model that allows memories to erode partially or disappear entirely.
An earlier version of this paper described the first three components and imposed a universal importance floor of 0.01, below which no memory could fall. All memories were immortal. This revision removes that floor and replaces it with a survival evaluation that treats memory persistence as something earned rather than guaranteed. The theoretical motivation for this change comes from two sources: the observation that biological memory involves genuine forgetting, not merely reduced accessibility; and the development of a geometric model of affective state (Riggleman, 2026b) that provides a principled basis for stickiness that the original paper lacked.
2. Related Work
2.1 Generative Agents
Park et al. (2023) describe a multi-agent simulation where LLM-based agents maintain memory streams, reflect on experiences, and plan future behavior. Their retrieval model scores memories on recency, importance, and relevance and takes a weighted sum. Potato’s memory architecture draws from this recency-importance framework but departs from it in two ways: importance is not static but decays over time according to a sliding window model, and access frequency feeds back into decay rate through a reconsolidation mechanism.
2.2 FadeMem and Access-Frequency Decay
FadeMem (Wei et al., 2026) is the nearest prior art to this paper, and closer than a citation for access-frequency decay alone would suggest. It proposes biologically inspired forgetting for agent memory: differential decay rates across a dual-layer memory hierarchy, with retention governed by adaptive exponential decay modulated by semantic relevance, access frequency, and temporal patterns. The overlap with the present architecture is substantial. Differential per-layer decay parallels the per-trace-type rates of Section 3.7; access-frequency-modulated exponential decay is the same mechanism as the sleep decay system of Section 3.4. FadeMem was posted in January 2026. The system described here was built independently beginning in March 2026, and its public development log provides timestamped evidence of that independence; the overlap is reported as what it is—convergence—rather than minimized.
The delta is affect. FadeMem modulates retention by relevance, frequency, and time, and by nothing else; no encoding-time emotional state appears anywhere in it. The present architecture adds: encoding-time emotional geometry as a decay modifier (Section 3.5); survival criteria that are structural rather than scalar—graph degree, identity relevance, contradiction links, centroid drift (Section 3.6); a four-tier forgetting taxonomy—full, fragment, ghost, true deletion—in which the tier a memory lands in is determined by its geometry (Section 3.7); and reinterpretation as versioned semantic traces rather than memory fusion (Section 3.8). The ablation of Section 4.2 is the empirical form of this delta: condition B is, in effect, FadeMem's retention model—frequency and recency without affect—and it produced survival outcomes that were a function of access schedule alone, deleting a moment of panic and a flat status exchange with equal indifference.
2.3 ACT-R Activation Theory
Anderson et al.’s (2004) ACT-R cognitive architecture models memory activation as a power-law decay function with access reinforcement. Each retrieval of a memory increases its base activation level, with the magnitude of reinforcement diminishing with repeated access. Potato’s reconsolidation model follows this principle: first recall matters most; the twentieth barely moves the needle.
2.4 Stratified Memory and the Polymorphic Memory Graph
The distinction between episodic and semantic memory (Tulving, 1972, 1985) and the further stratification of memory by cognitive role (Anderson et al., 2004) provide the theoretical basis for storing memories not as objective facts but as subjective experience, with agent-specific metadata reflecting the emotional and contextual state at encoding. Al-Kaddah (2026) names this approach the Polymorphic Memory Graph. Potato implements a simplified version using five vector collections with typed importance weights.
2.5 Constructive Memory and Reconsolidation
Schacter and Addis (2007) argue that episodic memory is constructive rather than reproductive: recall is not playback but reassembly. Nader and Hardt (2009) show that retrieved memories enter a labile state and must be reconsolidated, during which they can be modified. These findings motivate the present architecture’s treatment of reconsolidation as a meaning-changing event rather than simple reinforcement.
2.6 Geometric Affective State
Riggleman (2026b) describes a three-dimensional affective state space defined by valence, activation, and intensity, with a centroid model for computing the resultant emotional state from simultaneous inputs and a spread scalar measuring internal emotional disagreement. The present paper extends that geometric model to define memory stickiness: the resistance of a memory to decay as a function of its position in affective space at encoding time. This is a one-way dependency. The geometric model does not require the memory decay model; the memory decay model requires the geometric model to define stickiness.
3. Memory Architecture
The system is built on SQLite with the sqlite-vec extension for vector search, using 384-dimensional vectors via FastEmbed (all-MiniLM-L6-v2).
3.1 Memory as Trace Bundle
A memory is not a single record. It is a bundle of traces, each serving a different cognitive function and each subject to its own decay curve. This is the foundational change from the earlier version of this paper, which treated each memory as a monolithic object with a single importance value.
Each memory consists of four trace types:
Raw trace. The closest representation to the original experience. For a conversation, this is the verbatim exchange. For a visual episode, this is the image, latent, or perceptual residue. The raw trace is the residue of what happened. It is the most expensive to store and the first to decay.
Semantic trace. A compressed interpretation of what the memory meant at encoding time. This is a reading of the raw trace, not a replacement for it. It is historical: what the system believed the memory meant when it was formed. Semantic traces are cheaper to store and decay more slowly than raw traces.
Graph trace. Structured relations to other memories, identity nodes, contradiction links, and provenance records. The graph trace stores what this memory connects to and why. It is the cheapest to store and the most resistant to decay, because relational structure is the most information-dense representation of a memory’s role in the agent’s knowledge.
Life history. Metadata tracking access count, last accessed timestamp, stickiness score, reconsolidation history, decay state, and the emotional centroid at encoding. Life history does not decay. It persists as long as any other trace in the bundle survives.
This decomposition is modality-agnostic. The same four-trace structure applies whether the original experience was textual, visual, auditory, or any other modality. The decay logic does not care about format. It cares about function.
3.2 The Sliding Window Decay Model
Every conversation enters memory at importance 1.0. On the next exchange, all recent memories decay one step according to a fixed ten-step curve.
Step
Importance
0 (current)
1.00
1
0.80
2
0.60
3
0.45
4
0.30
5
0.20
6
0.12
7
0.07
8
0.03
9
0.01 (background)
Table 1. Sliding window importance decay curve.
This is not a technical optimization to save context window space, though it does that. It is a model of how attention works. Recent things are vivid. Older things fade unless something recalls them. In the earlier version of this paper, memories at background importance (0.01) represented a permanent floor. In this revision, 0.01 is no longer a floor but a waypoint. Memories can continue to decay below this threshold and, under conditions described in Section 3.6, can be partially or fully deleted.
Memory retrieval is importance-weighted: results are ranked by a blend of 80% semantic similarity and 20% importance. Decayed memories need a precise semantic match to surface against more important competitors.
The system stratifies memory types by cognitive role. Conversations are lived experience, entering at full importance and decaying naturally. Dreams are subconscious processing at sub-background importance (0.005), always labeled to prevent contamination of factual recall. Biases are self-knowledge at 0.03. Insights are confirmed understanding at 0.50. Explicit tagging and importance stratification prevents the agent from presenting its own pattern-matching as reality.
3.3 Reconsolidation
When a memory surfaces during conversation and is actually used, it gets stronger. The access timestamp and count are updated, and importance increases by a bounded amount that diminishes with repeated retrieval.
Memory Type
Reconsolidation Cap
Conversation
1.00
Insight
0.80
Bias
0.10
Dream
0.005
Table 2. Reconsolidation importance caps by memory type.
Topics the user brings up repeatedly become harder to forget. Topics discussed once decay on schedule. The agent does not choose what to remember—its memories choose themselves by being useful, relevant, and recalled.
Internal processing (dreaming, bias detection) does not trigger reconsolidation. Only external queries count as recall events. This prevents background processing from artificially inflating memory importance.
Reconsolidation also interacts with the geometric model described in Section 3.5. When a memory is reconsolidated, the emotional context of the recall event may differ from the emotional context at encoding. This produces a shift in the memory’s associated centroid. The magnitude and direction of centroid drift across reconsolidation events is itself informative: a memory whose emotional reading changes substantially over time is one the agent has not yet resolved. Increasing centroid drift is a signal to preserve the memory for further processing.
3.4 Sleep Decay and the Flashbulb Effect
Once per night during a 04:00 wake cycle, the system applies sleep decay to all conversation memories. The base decay rate is 0.85: an untouched memory retains 85% of its importance each night, modified by access history. The full nightly update rule is:
rate = min(0.85 + R + F + G + S, 0.98) R (recency) = 0.05 · (1 − h/24) if h < 24, else 0 (h = hours since last access) F (frequency) = 0.01 · min(access_count, 10) G (geometry) = 0.12 · max(activation_norm, sigma_distance_norm) (Section 3.5) S (spread) = 0.03 if spread_at_encoding > 0.5, else 0 (Section 3.5.1) importance ← importance · rate
Every retention figure in this paper is derivable from this rule. Table 3 tabulates its steady-state behavior for representative access patterns with the geometric terms at zero (G = S = 0):
Access Pattern
Nightly Decay Rate
7-Day Retention
Never accessed
0.85x
32%
Accessed twice, not recently
0.87x
38%
Accessed 4 times, not recently
0.89x
44%
Accessed 10+ times, recalled within 24 hours
0.98x (capped)
87%
Table 3. Access-aware sleep decay—the flashbulb effect. Each rate is the update rule evaluated at steady state for that access pattern (F = 0.01·min(count, 10); R = 0.05 at one hour since recall; capped at 0.98). Seven-day retention = rate⁷. These figures isolate decay; reconsolidation boosts from retrieval events (Section 3.3) are excluded.
After seven nights, a memory accessed ten or more times and recalled just before sleep retains 87% of its original importance (0.98⁷ ≈ 0.87). A memory never accessed retains 32% (0.85⁷ ≈ 0.32). In practice a frequently recalled memory retains more than the decay-only figure, because each retrieval also adds a reconsolidation boost; Table 3 reports the decay pathway alone so that every cell is reproducible from the stated rule. Combined with the sliding window, three decay forces operate at different timescales: intra-conversation attention decay, inter-day aging, and recall-based reinforcement.
In this revision, sleep decay also applies per-trace-type rates as described in Section 3.7. The overall importance value described here governs retrieval ranking. The per-trace rates govern which components of the memory bundle survive.
3.5 Geometric Stickiness
The earlier version of this paper used access frequency as the sole determinant of decay resistance. This is insufficient. A memory discussed once during a moment of intense fear should be stickier than a memory discussed three times during routine small talk. The missing variable is emotional state at encoding, and the geometric model of affective state (Riggleman, 2026b) provides the principled basis for incorporating it.
The affective state space is defined by three axes: valence (negative to positive), activation (low arousal to high arousal), and intensity (weak to strong). The agent maintains a baseline emotional state, sigma, representing its resting centroid in this space. Every memory is tagged at encoding with the agent’s emotional centroid and spread at the time of formation.
Two geometric quantities govern stickiness:
Activation. The activation coordinate of the emotional centroid at encoding is the primary stickiness driver. High activation corresponds to high arousal—the neurochemical mechanism (norepinephrine, amygdala-mediated consolidation) that cements memories in biological systems (McGaugh, 2004). Memories encoded during high-activation states resist decay directly, regardless of valence. A moment of panic and a moment of euphoria at the same activation level should be equally sticky. This is consistent with the arousal-based consolidation literature: it is arousal, not pleasantness, that predicts retention.
Distance from sigma. The Euclidean distance from the encoding centroid to the agent’s sigma provides a general stickiness measure that captures cases activation alone would miss. Deep grief, quiet devastation, and profound peace may sit at low activation but high intensity and extreme valence—points far from sigma in the full three-dimensional space. Distance from sigma flags these as emotionally remarkable without requiring high activation. This is the geometric mechanism: any memory encoded far from the agent’s emotional baseline is remarkable by definition, regardless of which axis or combination of axes pushed it away from sigma.
The two mechanisms are complementary but not independent. Activation is a component of the sigma-distance computation—it enters directly as one axis of the Euclidean distance, and again through derived intensity, which is computed from displacement in the valence/activation plane. The two signals are therefore correlated by construction. Computed over a uniform grid of the state space (valence, activation ∈ [−1, 1], intensity derived; activation normalized as clamp(a, 0, 1), arousal above neutral; sigma distance normalized by its maximum over the domain), the two signals correlate at r = −0.09 for the agent's baseline sigma, and the max() operator selects sigma distance for 67% of states and activation for the remaining 33%. The weak negative correlation reflects the geometry: states far below baseline arousal—deep calm, quiet grief—sit far from sigma while scoring zero on activation. Activation is best understood as a fast proxy for arousal-mediated consolidation; sigma distance is the general case, catching emotional remarkability at any arousal level. Taking the maximum is a cheap way to catch either. A memory that scores high on either signal resists decay. A memory that scores low on both has no geometric basis for survival and decays at the base rate.
The stickiness modifier applied during sleep decay is:
stickiness = max(activation_normalized, sigma_distance_normalized)
This is not a weighted sum. The dominant signal wins. High activation cements the memory through arousal. High sigma distance cements it through geometric weight. The memories that die are those where both values are low: flat, unremarkable moments near the agent’s emotional baseline.
Valence does not contribute to stickiness. It determines the character of the memory—whether the moment was pleasant or aversive—but not whether the memory survives. This is a deliberate design choice consistent with the consolidation literature and with the geometric model’s treatment of valence as a directional axis rather than a magnitude axis.
3.5.1 Spread as Unresolved Tension
The spread scalar from the centroid model (Riggleman, 2026b) measures the degree of internal emotional disagreement at the time of encoding. High spread means the agent was experiencing mixed or conflicting emotional inputs simultaneously. Memories with high spread at encoding decay more slowly because they represent unresolved states—the agent has not settled on a clean emotional reading of the experience. This unresolved quality is itself a reason to preserve the memory. It is still doing cognitive work.
If spread increases across reconsolidation events—that is, the agent becomes more conflicted about a memory over time rather than less—this is a strong preservation signal. The memory is actively generating new interpretive tension and should not be deleted.
3.6 Survival Evaluation and True Forgetting
The most significant change in this revision is the removal of the universal importance floor. In the earlier version, no memory could fall below 0.01. All memories were immortal. This revision replaces the floor with a survival evaluation that runs during nightly sleep decay.
When a memory’s overall importance drops below a threshold (the death line, set at 0.005), it enters candidacy for deletion. Deletion is not automatic. The system evaluates the memory against a set of survival criteria. If any criterion is met, the memory survives. If none are met, the memory proceeds through progressive erosion.
3.6.1 Survival Criteria
A memory below the death line survives if any of the following conditions hold:
• Graph degree above threshold: the memory has active connections to other surviving memories. Deleting it would orphan relational links.
• Identity relevance flag set: the memory touches self-knowledge, self-narrative, or the agent’s model of the user.
• Contradiction link exists: the memory is in tension with another memory. Deleting it would erase evidence of disagreement.
• Geometric stickiness above threshold: the memory’s activation or sigma distance at encoding exceeds the stickiness threshold defined in Section 3.5.
• High spread at encoding or increasing spread across reconsolidations: the memory represents unresolved emotional tension.
• High centroid drift: the memory’s emotional reading has shifted substantially across reconsolidation events, indicating ongoing cognitive processing.
• Cumulative access count above threshold: the memory was recalled multiple times during its lifetime, indicating historical significance even if not recently accessed.
3.6.2 Deletion Criteria
A memory below the death line is deleted if all of the following conditions hold:
• Low or zero access count.
• No identity relevance.
• No active graph connections to surviving memories.
• No contradiction links.
• Low geometric stickiness (low activation and low sigma distance at encoding).
• Low spread at encoding with no spread increase across reconsolidations.
• Sustained time below the death line (not a single bad night but extended irrelevance across multiple sleep cycles).
The key principle is that survival is earned, not guaranteed. A memory must have connections, identity relevance, emotional geometry, or demonstrated use to persist. An isolated, unaccessed, emotionally flat memory with no graph connections has no basis for continued existence. This is a more honest model than a universal floor that grants immortality to every record regardless of its role in the system.
3.7 Fragment Survival and Differential Trace Decay
Because a memory is a bundle of traces, forgetting does not have to be all-or-nothing. Different traces decay at different rates, producing partial forgetting that mirrors the phenomenology of biological memory: you forget the exact words but remember the gist; you forget what it meant but remember that something important happened; you forget everything except the connection to another memory.
Trace Type
Base Nightly Decay
Survival Bias
Raw (verbatim/image)
0.80x
Lowest—expensive, often redundant
Semantic (summary)
0.88x
Medium—compact, useful for retrieval
Graph (connections)
0.95x
Highest—cheap, most information-dense
Life history
No decay
Persists while any trace survives
Table 4. Per-trace-type nightly decay rates.
This produces four tiers of memory state:
Full survival. All traces intact. Importance above background. The memory is fully retrievable with original content.
Fragment survival. Raw trace deleted. Semantic and graph traces remain. The agent cannot reproduce the original exchange but remembers what it was about and how it connects to other knowledge.
Ghost survival. Raw and semantic traces deleted. Only the graph trace remains. The agent knows that a connection exists—“I know I talked to this person about something important”—but cannot recover what or why.
True deletion. All traces removed. Graph links severed. The memory is gone. No trace, no link, no residue. This is genuine forgetting.
The per-trace decay rates interact with the geometric stickiness modifier from Section 3.5. A memory with high stickiness has all its trace decay rates slowed proportionally. A memory with low stickiness has its raw trace decay accelerated, pushing it toward fragment survival faster. Over time, the memory store naturally shifts from a collection of detailed records to a web of thin connections with sparse anchoring content. This is consistent with the observation that long-term human memory is primarily relational and reconstructive rather than archival.
3.8 Reinterpretation
When a memory is reconsolidated, the system may generate a new semantic trace rather than simply boosting the existing one. This models the constructive nature of recall described by Schacter and Addis (2007): retrieval is not playback but reassembly, and the reassembled memory may carry new meaning.
Each reinterpretation creates a new interpretation node linked to the original memory bundle. The old interpretation is not overwritten. It becomes one reading among potentially many, tagged with the emotional centroid and context at the time of recall. This produces a versioned history:
interpretation_v1: childhood bedroom, bunkbeds, shared room (encoding context) interpretation_v2: strongest surviving anchor is bunkbeds, room mostly lost (5 years later) interpretation_v3: memory tied to security, crowding, sibling presence (current context)
This matters because new context can generate new insight without rewriting the original past. The record of what was originally believed is preserved alongside the record of what is currently believed. The core principle is: the raw trace is the residue; the semantic trace is a reading; readings can change; residue does not.
4. Observed Behavior
The sliding window model produces observable effects over multi-day use. Topics discussed repeatedly remain easily retrievable weeks later due to reconsolidation reinforcement and slower sleep decay. Topics raised once and not revisited fade to background importance and require specific semantic cues to surface. The agent does not volunteer these topics. They are present in storage but not in active context. This matches the design intent: what the agent remembers should track what the user actually brings up.
The geometric stickiness model and true forgetting system were validated in a controlled simulation rather than live deployment; production wiring of encoding geometry is partial. The experiments below test the architecture's two new claims directly.
4.1 Ablation Design
The revision's central new claim is that geometric stickiness adds something access frequency alone does not. This was tested with a controlled ablation. A fixed memory set—six encoding-affect profiles (flat-at-baseline; mildly neutral; high-arousal positive; high-arousal negative; low-arousal-far, modeling deep grief; and high-spread) crossed with four access schedules mirroring Table 3's rows (never accessed; accessed twice early; accessed four times spread out; accessed every night), two replicates per cell, 48 memories per condition, all entering at importance 1.0—was seeded into two identical databases. Condition A ran nightly sleep decay with geometric stickiness enabled. Condition B ran the same nights with the geometric terms removed from both the decay modifier and the survival evaluation, leaving access frequency as the only decay resistance. Nights were simulated on a virtual clock (24 hours per night, 40 nights), so recency bonuses behaved exactly as in deployment. The harness is fully deterministic: replicates within a cell produce identical trajectories and serve as a determinism check, not statistical power. The effective sample is the 24 unique profile-by-schedule cells, and every comparison reported below is structural—a property of the update rule—rather than statistical. The harness and per-night importance traces are retained alongside the codebase (scripts/ablation_stickiness.py; results/ablation_stickiness/).
4.2 Ablation Results
The two conditions produced different surviving sets, not merely different decay speeds. After 40 nights, condition B had deleted all twelve never-accessed memories and eroded all twelve accessed-early memories to ghosts, regardless of emotional geometry: a moment of panic, a moment of euphoria, and a flat status exchange met identical fates when their access schedules matched. Condition B's outcomes were a function of schedule alone. Condition A preserved every emotionally remarkable memory with full traces—high-arousal memories at importance 0.18 after 40 unaccessed nights, deep grief at 0.13—while the flat-at-baseline unaccessed memory eroded to a fragment. All 48 memory pairs diverged between conditions.
Three specific predictions of Section 3.5 held. First, valence independence: the high-arousal positive and high-arousal negative profiles produced identical trajectories in condition A (importance 0.4124 after 40 nights on the accessed-early schedule, in both), confirming that arousal, not pleasantness, drives stickiness in the implementation as specified. Second, the sigma-distance mechanism caught the case activation misses: the low-arousal-far profile (activation below neutral, so its normalized activation contributes zero) survived 40 unaccessed nights at importance 0.125 on sigma distance alone; in condition B it was deleted. Third, erosion produced the predicted tiered outcomes—full, fragment, ghost, and true deletion all occurred, and which tier a memory landed in depended on its geometry in condition A and only on its schedule in condition B.
4.3 Measured Retention Against Table 3
Condition B doubles as a validation of Table 3, since with geometry disabled the decay rule reduces to the access-modified rates the table describes. Measured seven-night mean retention against the table's steady-state predictions: never accessed, 0.321 (predicted 0.32); accessed twice early, 0.415 (predicted 0.38); accessed on a spread schedule, 0.359 (predicted 0.44); accessed every night, 0.637 (predicted 0.87). The never-accessed row reproduces the table exactly. The others deviate in both directions for a mechanical reason the steady-state idealization hides: the frequency term scales with accumulated access count, so a memory does not decay at its steady-state rate until its accesses have actually happened. The accessed-early memories beat prediction because they also earned the recency bonus on their two access nights; the spread-schedule memories fell short because only one of their four accesses occurs within the first seven nights; the nightly-accessed memories fell short because the rate reaches its cap only after the access count has accumulated. Table 3 describes a memory that already has its access history; the measured curve describes one acquiring it. Both are consequences of the same update rule, and the gap is reported rather than reconciled because it is informative about how quickly the flashbulb effect engages.
4.4 A Defect Found by the Ablation
The first ablation run falsified the true forgetting mechanism as then implemented. No memory in either condition eroded or died across 40 nights, because the codebase retained a legacy importance floor of 0.01 from the earlier version of this architecture—above the 0.005 death line—making the survival evaluation unreachable. The revision's prose claimed the floor had been removed; the code disagreed. The floor has now been removed for non-identity memories (identity nodes retain their protected floor), and the results above are from the corrected run. This is reported because it could have come out the other way, and initially did: an architecture document is not evidence, and the gap between specified and implemented behavior in this case silently disabled the revision's most significant feature.
5. Limitations and Future Work
The ablation is a controlled simulation with synthetic memories and deterministic schedules, not a study of live use. All sliding window and reconsolidation observations are from a single-user deployment on a single machine. Live encoding of the geometric quantities is only partially wired into production code paths, so deployed behavior has not yet accumulated organic data against these predictions.
Future work should include: formal evaluation metrics for memory decay behavior relative to a static importance baseline; comparative studies against stateless LLM baselines on multi-session tasks; quantitative analysis of reconsolidation boost rates against ACT-R literature; replication of the stickiness ablation on organic memories from live deployment rather than synthetic seeded sets; evaluation on the benchmarks FadeMem reports against (Multi-Session Chat, LoCoMo, LTI-Bench), which would make the comparison of Section 2.2 quantitative; calibration of the death line threshold, survival criteria weights, and per-trace decay rates through longitudinal deployment; and investigation of whether spread trajectory (increasing vs. decreasing spread across reconsolidation events) predicts subjective memory significance as rated by users.
The interaction between geometric stickiness and distributed cognition architectures—whether modeled as a Society of Mind (Minsky, 1986), global workspace (Baars, 1988), or the Parliament of Mind pattern used in Potato (Riggleman, 2026a)—is an open question. If different cognitive modules weight memories differently, the survival evaluation may need to be module-aware rather than centralized.
6. Conclusion
The sliding window decay model, reconsolidation mechanism, nightly sleep decay system, geometric stickiness model, and true forgetting architecture together produce a memory system qualitatively different from a static context window. What the agent remembers is determined by what the user engages with, what gets recalled, how recently, and what emotional geometry was present at encoding. Memory that matters—that was formed during remarkable emotional states, that earned connections to other knowledge, that was recalled and found useful—resists forgetting. Memory that does not get used, that was formed near the emotional baseline, that connects to nothing and serves no identity function, erodes and eventually disappears.
This is a functional approximation of the same principle that governs biological memory: salience and use determine what survives. The addition of geometric stickiness derived from a three-dimensional affective state space makes this approximation more principled than access frequency alone. The addition of true forgetting makes it more honest than a universal floor that prevents any memory from being lost.
The strongest theoretical claim this architecture supports is that memory survival is not a property of the memory itself. It is a property of the memory’s relationship to the rest of the system: its emotional geometry, its graph connections, its identity relevance, its use history. An isolated, flat, unused memory has no reason to exist. A connected, emotionally charged, repeatedly recalled memory resists forgetting even without recent access. The system does not decide what to remember. The memories decide for themselves, through the structure they accumulate and the geometric weight they carry.
Acknowledgments
The memory stratification draws from the episodic-semantic distinction (Tulving, 1972, 1985) and the role-stratified memory systems of ACT-R (Anderson et al., 2004), with the naming of the “Polymorphic Memory Graph” due to Al-Kaddah (2026). The sliding window decay model and ten-step importance curve are original to the author, Brian Riggleman. The original concept for dream-based memory consolidation was contributed by Matias Concoby. The geometric stickiness model derives from the centroid model of emotional experience (Riggleman, 2026b).
References
Al-Kaddah, S. (2026). Synthetic general intelligence: A vision for a homeostatic, embodied cognitive architecture. Zenodo. https://doi.org/10.5281/zenodo.19034990
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), 1036–1060. https://doi.org/10.1037/0033-295X.111.4.1036
Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Brown, R., & Kulik, J. (1977). Flashbulb memories. Cognition, 5(1), 73–99. https://doi.org/10.1016/0010-0277(77)90018-X
Dudai, Y. (2004). The neurobiology of consolidations, or, how stable is the engram? Annual Review of Psychology, 55, 51–86. https://doi.org/10.1146/annurev.psych.55.090902.142050
Wei, L., Peng, X., Dong, X., Xie, N., & Wang, B. (2026). FadeMem: Biologically-inspired forgetting for efficient agent memory. arXiv preprint arXiv:2601.18642. https://arxiv.org/abs/2601.18642
McGaugh, J. L. (2004). The amygdala modulates the consolidation of memories of emotionally arousing experiences. Annual Review of Neuroscience, 27, 1–28. https://doi.org/10.1146/annurev.neuro.27.070203.144157
Nader, K., & Hardt, O. (2009). A single standard for memory: The case for reconsolidation. Nature Reviews Neuroscience, 10(3), 224–234. https://doi.org/10.1038/nrn2590
Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442. https://doi.org/10.48550/arXiv.2304.03442
Riggleman, B. (2026a). Potato: A Persistent On-device Temporal Agent with Tunable Ontology. Clawd Daily. https://clawddaily.com/potato
Riggleman, B. (2026b). The centroid model of emotional experience in a persistent embodied agent. Clawd Daily. https://clawddaily.com/papers/centroid-model
Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: Remembering the past and imagining the future. Philosophical Transactions of the Royal Society B, 362(1481), 773–786. https://doi.org/10.1098/rstb.2007.2087
Minsky, M. (1986). The Society of Mind. Simon & Schuster.
Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of Memory (pp. 381–403). Academic Press.
Tulving, E. (1985). Memory and consciousness. Canadian Psychology, 26(1), 1–12.