Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs

Motivation

• Existing systems treat personal facts uniformly, creating a lifecycle-management problem because preferences, identity facts, and logistics require different persistence and replacement dynamics.Prior systems provide varied memory infrastructures but do not distinguish current from superseded preferences, expired logistics, or approaching deadlines, while existing benchmarks do not measure these distinctions.Source: §1 · p. 2

Contribution

• Fortunate Recall classifies extracted facts into 10+1 behavioral categories and applies deterministic category-specific decay, supersession, event-time validity, and retrieval-routing policies.Retrieval uses a single closed-form log-score without a learned lifecycle mediator, with median 47 μs latency and empirically linear O(k) scaling.Source: §1 · p. 2

Contributions

• 76.9% on LifecycleBench and 75.2% on full LongMemEval-S under the Wu et al. protocol establish FR-Bank’s benchmark performance.FR-Bank is validated across seven configurations and two generator families using both answered-query and all-query denominators.Source: §1 · p. 2
• The paper contributes a behavioral ontology, a lifecycle non-identifiability theorem, LifecycleBench, FR-Bank, and attribution and transfer studies, releasing the ontology, benchmark, code, and supporting artifacts.LifecycleBench contains 516 temporal-disambiguation questions over 40 personas with nine attack vectors; studies include untyped ablation, BEAM transfer, granularity and metadata-noise sweeps, and judge audits.Source: §1 · p. 2

Related-work gaps

• Existing systems lack combinations of structured lifecycle policies, differential decay, expiry or retraction, event-time validity, and forgetting mechanisms.Mem0 [19] and A-MEM [4] lack structured lifecycle; Memory-R1 [10] uses flat ADD/UPDATE/DELETE operations; MemoryBank [5] applies uniform decay; MemGPT [8], APEX-MEM [17], and Zep/Graphiti [2] leave key lifecycle decisions unresolved.Source: §2 · p. 2

Fortunate Recall architecture

• Fortunate Recall separates LLM extraction from an inspectable deterministic lifecycle layer using behavioral category, slot key, lifecycle state, event-time, ingest time, and confidence.Its bi-temporal representation distinguishes event-time from transaction-time, and lifecycle decisions are computed as named score terms at millisecond latency.Source: §2 · p. 2, Figure 1 · p. 3

Benchmarks

• LifecycleBench targets a gap in evaluation by making every question depend on lifecycle state, unlike LongMemEval and LoCoMo [13], which omit current-versus-outdated-state discrimination.BEAM [14] covers ten memory abilities and includes lifecycle-sensitive tasks, but LifecycleBench is positioned as a dense lifecycle stress test rather than the first temporal benchmark.Source: §2 · pp. 2–3

Distinction from cognitive typologies

• Behavioral domains determine appropriate lifecycle dynamics better than cognitive memory types: chronic illness and lunch orders may both be semantic but require different treatment.MemoryOS and MIRIX [11] classify memory form, whereas Fortunate Recall distinguishes domains such as Health & Wellbeing and Logistical Context.Source: §2 · pp. 2–3

3 Fortunate Recall

• The architecture combines an infrastructure layer using bi-temporal validity and hybrid retrieval with a lifecycle policy layer providing behavioral dynamics.This separation makes lifecycle management composable across retrieval substrates.Source: §3 · p. 3

3.1 The Behavioral Ontology

• Fortunate Recall classifies each fact into a 10+1 behavioral ontology whose category determines decay, supersession, expiry, and retrieval routing.The ontology is organized by temporal behavior rather than cognitive form; Table 1 defines category half-lives and lifecycle dynamics.Source: §3.1 · p. 3, Table 1 · p. 4

3.2 Classify the Fact, Not the Entity

• Fact-level classification avoids confusing conversational framing or shared entities with the information being memorized, achieving κ= +0.67 agreement with full session context.For example, changing a surname is Identity, while catching bass is Hobbies, regardless of the surrounding utterance or entity.Source: §3.2 · p. 4

3.3 Lifecycle Mechanisms

• Lifecycle mechanisms combine differential decay, confidence-weighted slot supersession, event-time validity, anticipatory activation, and category-aware retrieval routing.Decay rates span 5.3×; contradictions can preserve both values when confidence is low, deadlines intensify near events, and category-forced retrieval returns up to 20 edges.Source: §3.3 · p. 4

3.4 Core Mechanism versus Implementation Engineering

• FR distinguishes core lifecycle mechanisms from implementation engineering by measuring the difference between configurations that retain or remove the classification layer.Table 2 identifies the separable mechanism and engineering components, while the untyped lifecycle baseline isolates classification’s contribution.Source: §3.4 · p. 5, Table 2 · p. 5

3.5 The LLM Boundary

• LLMs handle ingestion and candidate selection, while deterministic functions execute decay, supersession filtering, event-time checks, routing, and scoring in 47 μs median at k=20.The deterministic layer scales empirically as O(k), with k denoting scored candidate edges after distillation.Source: §3.5 · p. 5

3.6 FR-Bank: Infrastructure-Independent Implementation

• FR-Bank validates substrate independence as a standalone flat embedding-store stack that replaces Graphiti while preserving the Graphiti lifecycle parameters.It merges cosine, BM25, and category-forced candidates, distills to top-20, and applies the full lifecycle stack without graph storage, entity resolution, or deduplication.Source: §3.6 · p. 5

• FR’s deterministic lifecycle score combines semantic similarity with category priors and decay, event-time validity, slot-local supersession, and query-conditioned lifecycle compatibility.The generative model motivates the feature set, while the log-score makes the resulting lifecycle policies explicit; it is a modeling assumption rather than a runtime inference claim.Source: §4 · p. 6
• Theorem 1 shows that similarity and recency alone cannot identify lifecycle-aware rankings, imposing a binary error floor of P_e≥1/2 under equiprobable histories.Histories can share identical (σ, Δt) profiles while differing in hidden supersession structure, so flat scorers fail on at least one.Source: §4 · p. 6
• The metadata basis (c, κ, ξ, h) is sufficient, minimal, and individually necessary because its components distinguish category survival, slot competition, lifecycle state, and event-time direction.Removing any component respectively conflates preservation with suppression, independent with competing facts, current with historical queries, or upcoming with expired obligations; related flat systems have worst-case error ≥1/2.Source: §4 · p. 6
• 99.8% of 1,566 preservation–suppression cross-pairs are jointly infeasible under a global α, while category-specific blending restores feasibility without benchmark-fitted rates.Ten of eleven categories tolerate the full 0.25×–4× multiplier sweep at retrieval Jaccard ≥0.95, and a 20/20 persona holdout reports 77.0% vs. 76.8%.Source: §4 · p. 6
• Stale facts materially alter outcomes: clean contexts pass at 75.8% versus 6.0% when contaminated, predicting a 13.3pp FR-Graphiti–Mem0 advantage versus the observed 12pp.The decomposition treats staleness as outcome-flipping contamination rather than harmless noise, with π=0.08 for FR-Graphiti and π=0.27 for Mem0.Source: §4 · p. 6
• The theory does not establish that the 10+1 partition, decay rates, or pipeline are optimal, and baselines carrying lifecycle metadata fall outside the impossibility classes by construction.The paper therefore tests partition importance empirically; its stated prediction is that such baselines lose calibration rather than retrieval correctness.Source: §4 · pp. 6–7

Benchmark results

• FR-Bank leads AV1, AV3, AV4, and AV5 outright, ties AV9, while MemoryOS leads AV6 at 73% versus 64% and FR-Graphiti leads AV2 and AV8.Memory-R1 leads none; FR-Bank scores only 5% retrieval pass on AV7, which tests selective forgetting after explicit retraction.Source: §5 · p. 7

Benchmark design

• 516 questions span 40 personas, approximately 35 multi-session conversations, 18 simulated months, and nine attack vectors covering supersession, expiry, contradiction, retraction, aggregation, and numeric preservation.Structured YAML ground truth specifies supersession events, expiry dates, retraction language, and ambiguity-resolution rules.Source: §5 · p. 7

Evaluation protocol

• The benchmark uses synthetic conversations because temporal ground truth is otherwise unobservable, with all systems evaluated identically and transfer and co-design analyses reported separately.The attack vectors were defined from preliminary Mem0 and Zep/Graphiti failure modes before FR mechanisms were designed; §6.6 reports transfer to an externally constructed benchmark.Source: §5 · p. 7

6.1 LongMemEval-S Under the Wu et al. Protocol

• 75.2%±0.70pp pass@10 on full LongMemEval-S establishes FR-Bank above RMM’s 70.4% under the canonical Wu et al. judge protocol.The matched-subset lifecycle effect is aggregate-neutral (+2.2pp), with knowledge-update performance decreasing 6.4pp because the judge rewards superseded facts appearing alongside updated answers.Source: §6.1 · p. 8

6.2 LifecycleBench Results

• 76.9% pass rate on LifecycleBench puts FR-Bank 16pp above Mem0 with 1.76× lower staleness.Full lifecycle adds +12pp over uniform decay; category-aware routing contributes +6pp, and 94.5% pass when the correct edge appears in top-10 indicates retrieval recall is the bottleneck.Source: §6.2 · p. 8
• FR-Bank reaches 37.5% end-to-end correctness on retraction cases, exceeding MemoryOS’s 7.5% despite MemoryOS’s higher 57% retrieval pass rate.Lifecycle state masking excludes retracted facts, enabling abstention rather than confabulation; generic summaries can inflate retrieval scores while remaining uninformative.Source: §6.2 · p. 8, §6.4 · p. 9

6.4 End-to-End Response Quality

• FR-Bank’s confabulation ranking remains ahead of Mem0 across generators, with 22.4% versus 45.1% over answered queries and 13.0% versus 32.2% over all queries.Correct-over-total is 31.2% versus 18.6%, and safe response rate is 73.3% versus 47.3%; Kimi K2.5 preserves the core ranking.Source: §6 · p. 7, §6.4 · p. 9

6.5 Attribution: What the Ontology Contributes

• The untyped lifecycle arm is correctness-equivalent to the full stack (Δ = −1.7pp, 95% CI [−6.0, +2.7]), while typing halves confabulation from 24.2% to 12.0%.Generic slot-key supersession, event-time expiry, and retraction masking carry correctness; behavioral typing increases calibration through safer abstention.Source: §6.5 · p. 9

6.6 External Transfer: BEAM

• On BEAM, FR-full is correct on 131 of 280 (46.8%) versus 92 (32.9%) for Mem0, with the ontology’s gain concentrated in contradiction resolution.Contradiction resolution is 29/70 for FR-full versus 19/70 for the collapsed ontology and 5/70 for Mem0; performance peaks near k=7 and then saturates.Source: §6.6 · p. 10, §6.5 · p. 10
• BEAM reproduces the correctness decomposition but not the confabulation decomposition, while slot-key fragility can make lifecycle-off knowledge-update performance exceed both typed arms.Confabulation is 108 →84 →72 in the primary run, whereas lifecycle-off scores 135/280 overall and 43 versus 33 on knowledge update.Source: §6.6 · p. 10

6.7 Robustness and Validation of Our Own Judge

• Judge audits raise absolute correct and safe rates 5–7pp but leave every cross-system ranking unchanged, making comparisons more robust than absolute rates.The pinned judge agrees with the cross-family panel at Fleiss κ=0.715 versus panel self-agreement κ=0.834; metadata noise is most damaging for slot-key corruption.Source: §6.7 · p. 11

7 Discussion and Conclusion

• FR-Bank answers 57.9% of queries, but its abstention benefit depends on deployments that need not always answer.The extra abstentions come from would-be wrong answers, creating a real correctness–abstention trade-off.Source: §7 · p. 11
• The work lacks human validation, maximally tuned competitor comparisons, multi-framework or multi-agent evaluation, and fully independent benchmark construction.LifecycleBench was co-designed from observed baseline failures; BEAM reduces but does not eliminate that risk, and its runs are not commensurable with LifecycleBench tables.Source: §7 · p. 11
• 77.0% versus 76.8% on a 20/20 persona holdout supports calibration generalization, while soft supersession fails under high-commitment Kimi AV9 generators.The seven configurations span five architectural paradigms, and rate calibration is constraint-driven rather than metric-optimized.Source: §7 · p. 11

Privacy, consent, and safety.

• Memory systems retaining personal facts raise unresolved concerns about consent, verifiable deletion, prompt-injection defenses, and cultural generalization of decay assumptions.These privacy and safety issues are discussed as beyond the scope addressed by this work.Source: Privacy, consent, and safety · p. 11

7 Discussion and Conclusion

• 22.4% versus Mem0’s 45.1% confabulation over answered queries and 13.0% versus 32.2% over all queries, while two implementations achieve indistinguishable LifecycleBench rates (p=0.10).The paper attributes correctness gains to generic metadata, while the ontology enables per-category calibration, halves confabulation, and provides a replicated contradiction-resolution gain saturating near seven policy clusters.Source: Privacy, consent, and safety · p. 11

A.1 Model 1 and Its Deterministic Consequence

• A query-conditioned compatibility mask deterministically filters lifecycle states, while FR applies lifecycle rules rather than posterior probabilities.The mask can admit active or superseded facts for historical queries and exclude retracted facts for deletion-sensitive queries.Source: §A.1 · p. 14

A.2.1 Exact non-identifiability

• Feature-restricted scorers using only semantic similarity and elapsed time can rank identical observations differently across histories, causing at least 1/2 error under equiprobable histories.The histories differ in supersession structure: one requires preferring the later replacement, while the other leaves both facts active and favors the more aligned fact.Source: §A.2.1 · p. 14

A.2.2 Deterministic ε-relaxation

• If 2Lε < Δmin, a Lipschitz feature-restricted scorer produces the same ranking for every intent and conveys zero intent information.The condition preserves every adjacent score relation, so the full ranking is constant across the query-intent ensemble.Source: §A.2.2 · p. 15

A.2.3 A looser stochastic entropy bound

• The entropy bound is only a weaker corollary requiring the additional comparison-flip model, whereas the deterministic margin theorem requires no such stochastic assumption.The supplied analysis explicitly treats the deterministic result as the exact operational statement and the entropy formulation as a softer framing.Source: §A.2.3 · p. 15, §A.2.4 · p. 16

A.2.4 Measured values, two regimes, and the binary case

• In the long-term regime, activations collapse to approximately 0, making the scalar family infeasible; FR instead uses supersession filtering, event-time expiry, and category-forced retrieval.Measured medians were |ε| = 0.060, |Δa| ≈ 3.3 × 10−10, and αrequired = 0.9998; the deterministic argument retains a 1/2 error floor.Source: §A.2.4 · p. 16

A.3 Proof of Theorem 2 and A.4 Proof of Minimality (Theorem 2′)

• The metadata tuple (category, slot key, lifecycle state, event time), together with similarity, elapsed time, history, and query, is sufficient and minimal for lifecycle-aware scoring.Every strict subset of {c, κ, ξ, h}, even with similarity and elapsed time, admits a task where deterministic scorers incur at least 1/2 error while the full basis can achieve zero.Source: §A.3 · p. 16, §A.4 · p. 17

A.5 The Behavioral Ontology as an Empirical Variance-Minimizing Partition

• The behavioral partition reduces within-category policy variance to 3.60 × 10−3, versus 2.09 × 10−2 for cognitive and 4.69 × 10−2 for uniform partitions.It achieves the optimal 11-group floor; the cognitive 3-partition is 4.2× worse than the optimal 3-partition because it merges Obligations and Logistical Context despite divergent lifecycle policies.Source: §A.5 · pp. 18–19, Table 5 · p. 19

A.6 Partition Separation: The Behavioral Variance Gap

• A weighted pairwise variance identity attributes coarsening excess to between-category policy differences, while within-category noise contributes an invariant floor across partitions.The analysis uses this identity to explain why merging behavioral categories increases variance under the category-mean model.Source: §A.6 · p. 19
• The measured partition gap is 36.9%, exceeding the 10.1% analytical lower bound because soft membership weights amplify cross-category spread.The pooled variance is dominated by α_c rather than λ_c, making the separation architecturally meaningful rather than a free-hyperparameter artifact.Source: §A.6 · p. 20

A.7.1 Removing category labels 𝑐

• Removing category labels makes one scorer unable to preserve older Financial facts while suppressing newer Logistical conflicts with identical visible features.The measured α bounds conflict: preservation requires α≤0.065 in the median case, whereas suppression requires α≥0.138 on FR-Graphiti.Source: §A.7.1 · pp. 20–21

A.7.2 Removing slot keys 𝜅

• Without slot keys, identical edge tuples cannot distinguish supersession within a shared slot from coexistence across different slots, forcing over- or under-deletion.Thus slot keys are necessary for correct deletion actions.Source: §A.7.2 · p. 21

A.7.3 Removing lifecycle state 𝜉

• Without lifecycle state, a scorer cannot reverse rankings between current-preference and change-history queries, so one intent must fail.The witness uses two same-slot preference edges with identical visible metadata but opposite query-conditioned requirements.Source: §A.7.3 · p. 21

A.8 Proof of Event-Time Oblivion (Theorem 3′)

• Without event-time anchors, deterministic retrieval cannot distinguish upcoming from expired obligations and incurs at least 1/2 error on a suitable query distribution.This impossibility applies to any scorer using age alone, including monotone age-only kernels, and requires information equivalent to the event-time sign.Source: §A.7.4 · p. 21, §A.8 · p. 22

A.9 Proof of Retraction Oblivion (Theorem 3′′)

• Binary-flat systems have Bayes error exactly 1/2 for mixed current-state and change-aware retraction queries, whereas retained-but-masked retraction state realizes both intents.Retraction therefore cannot be represented by deletion or ordinary active retention alone.Source: §A.9 · pp. 22–23

A.10.2 Global-α infeasibility as a one-dimensional slice

• 99.8% of FR-Bank cross-pair constraints are infeasible under a global α, while all 45 populated cross-category cells lack a feasible shared value.The 11 × 11 polytope has zero feasible cells, demonstrating that one-dimensional parameterization cannot satisfy heterogeneous preservation and suppression requirements.Source: §A.10 · p. 23, §A.10.2 · p. 24

A.11 Dimensional Necessity of Category-Specific Parameters

• A dimensionality claim shows that any affine parameter family of dimension below K is generically unable to realize independently perturbed category-margin targets.The argument identifies category-specific {α_c, λ_c, F_c} as the coordinates spanning the category-margin space, matching the empirical FR-Bank witness.Source: §A.11 · p. 24

A.12 Staleness Decomposition: Derivation and Numerical Predictions

• The exact staleness decomposition is P_S(pass)=(1−π_S)p0+π_Sp1, separating contamination prevalence from clean- and contaminated-context pass rates.The across-system gap therefore depends on both contamination differences and calibration of the conditional pass rates.Source: §A.12 · p. 24
• 13.3pp of FR-Graphiti’s pass advantage over Mem0 is predicted by contamination reduction alone, close to the observed 12pp retrieval gap.The calculation uses clean-vs-contaminated pass rates p0=0.758 and p1=0.060, with staleness rates 0.08 and 0.27.Source: §A.12 · p. 25
• FR’s staleness reduction corresponds to lower confabulation, with 22.4% versus Mem0’s 45.1% over answered queries.The confabulation analogue is P_S(confab)=(1−π_S)c0+π_Sc1, so contamination differences map directly to confabulation differences when conditional rates are fixed.Source: §A.12 · p. 25
• 76.9% pass rate for FR-Bank reflects both low staleness (π=0.155) and high clean-context recall, because staleness alone cannot predict performance.FR-Bank and Memory-R1 have nearly identical staleness rates (0.155 vs. 0.153) but pass rates differing by 10.0pp, falsifying system-invariant clean and contaminated-context rates.Source: §A.12 · p. 25

A.13 Anticipatory Activation and Why Age Alone Cannot Express It

• No monotone nonincreasing function of edge age can represent anticipatory activation, because relevance must increase before an event and decrease after it.FR’s event-time kernel implements this shape: relevance rises as the event approaches and decays after expiry.Source: §A.13 · p. 25

A.14 Proof of Lemma 1 (Harmonic-Mean Decay)

• Soft category memberships make harmonic-mean decay operationally important: 66.8% of FR-Bank’s 12,968 entries have primary-category weight below 0.90.The mean entry spans 2.0 categories with weight above 0.05, so mixed-category decay affects most of the corpus rather than a theoretical edge case.Source: §A.14 · p. 26

A.15 Final Interpretation

• The formal results identify (c, κ, ξ, h) as a necessary metadata basis and interpret FR as a deterministic lifecycle policy, not runtime Bayesian inference.The theory also treats calibration as a convex feasibility problem and identifies clean-context recall as orthogonal to staleness filtering.Source: §A.15 · p. 26

B.1.1 Classifier Agreement Audit

• 70.7% majority-vote agreement (κ=+0.673) on 200 held-out facts supports the classifier, with strongest reliability in categories having divergent lifecycle policies.The audit used three independent judges with full session context; macro-F1 was 0.595, weighted F1 was 0.647, and accuracy was 66.5%.Source: Table 7 · p. 28, Table 8 · p. 28, Table 10 · p. 29

B.1.2 Downstream Sensitivity to Category Perturbation

• At 50% label perturbation within confused category pairs, top-10 retrieval retained 0.905 Jaccard overlap and Kendall τ remained 0.897, showing limited downstream sensitivity.At 10% perturbation, Jaccard overlap was 0.974 ± 0.002 with 0.29 displaced edges per query.Source: §B.1.2 · p. 29, Table 11 · p. 30

B.2 Ontology-Aware Session Initialization (Warm Start Protocol)

• Warm sessions traverse the active frontier, while evolving sessions expand context asynchronously after sustained cosine-distance drift from the rolling query-category vector.Cold sessions skip graph traversal, and the drift trigger requires exceeding threshold δ for N consecutive turns.Source: §B.2 · p. 30

B.3 Per-User Parameter Evolution

• The system supports future per-user decay adaptation from access frequency, temporal gaps, and supersession rates, but empirical validation of that loop is deferred.Frequent hobby mentions would lower Hobbies decay, whereas rapidly changing preferences would increase Preferences decay.Source: §B.3 · p. 30

B.4 Interpretability and User Control; B.5 Safety and Robustness

• Human-readable categories and lifecycle policies make memory inspectable and editable, while safety mechanisms quarantine instructional injections, flag sensitive inconsistencies, expose ambiguous supersessions, and audit deletion.Health and Financial facts can require reconfirmation, and low-confidence supersessions surface both candidate facts.Source: §B.4 · p. 30, §B.5 · p. 31

B.6 Emotional State as Signal

• Transient emotion is modeled as a fast-decaying activation signal on the relevant category rather than as a persistent memory category.Frustration about a work deadline routes to Obligations with an emotional-loading modifier.Source: §B.6 · p. 31

C LifecycleBench: Extended Methodology; C.1 Persona Design; C.2 On Synthetic Generation

• LifecycleBench uses fully specified temporal personas spanning 35 sessions and 18 simulated months, with explicit supersession, expiry, retraction, contradiction, and ambiguity ground truth for lifecycle-state evaluation.Its controlled synthetic design enables evaluation of lifecycle state management beyond retrieval recall alone.Source: §C.1 · p. 31, §C.2 · p. 32

C.3 Persona Examples; C.4 Scale and Coverage; C.5 Evaluation Metrics

• The corpus contains 18,936 entries across 12,460 slot keys, while evaluation prioritizes pass rate and staleness over MRR because retrieved facts receive near-equal attention.The benchmark reports top-K answer support and outdated-edge penalties using a Claude Sonnet judge against structured ground truth; four configurations vary decay type and routing.Source: §C.4 · p. 32, §C.5 · p. 32

C.6 Benchmark Independence Analysis

• Table 13 illustrates coverage across five personas for AV1, AV2, AV4, AV7, and AV9, spanning supersession, logistics validity, multi-version facts, retraction, and ambiguity.Source: Table 13 · p. 33
• LifecycleBench covers nine temporal-state failure modes intrinsic to persistent personal memory, including replacement, expiration, retraction, partial updates, contradiction, aggregation, and numeric-value loss.These attack vectors are presented as properties any independently constructed lifecycle benchmark would need to cover.Source: §C.6 · p. 33
• FR does not dominate LifecycleBench: MemoryOS leads AV6 at 73% versus FR-Bank’s 64%, while FR-Graphiti leads AV2 and AV8.FR-Bank records 5% retrieval pass on AV7 and near-zero end-to-end correct rates on AV5 and AV6, while FR-Bank leads AV3 at 97%.Source: §C.6 · pp. 33–34
• Table 14 situates LifecycleBench’s scale relative to LongMemEval, providing the benchmark-size comparison used in the independence analysis.Source: Table 14 · p. 34
• Cross-validation on LongMemEval-S shows lifecycle is aggregate-neutral on the 317-question matched subset (+2.2pp net), with costs concentrated where stale facts remain accepted.Knowledge-update falls −6.4pp because the Wu et al. judge marks answers correct when stale facts accompany the updated answer.Source: §C.6 · p. 34
• The behavioral ontology outperforms the cognitive partition on 7 of 9 attack vectors under the same pipeline, indicating LifecycleBench discriminates among ontology approaches.The comparison is reported as an ontology ablation in Table 4.Source: §C.6 · p. 34
• The authors acknowledge that co-design risk cannot be fully eliminated without independent replication and release the benchmark materials and evaluation code for identical-condition comparisons.Released materials include personas, conversations, questions, ground truth, and evaluation code.Source: §C.6 · pp. 32, 34

D.1 LongMemEval Detailed Results

• 75.2% mean LongMemEval-S accuracy across ten identical reruns, with every run at least 74.2% and a maximum of 76.6%.The evaluation uses Wu et al. (ICLR 2025)’s 500-question canonical setup with temperature 0, identical retrieval, and top-10 facts.Source: §D.1 · p. 34, Table 15 · p. 34

D.2 Lifecycle Contribution Decomposition

• +2.2pp lifecycle gain on the 317-question matched subset, concentrated in knowledge-update and multi-session questions requiring temporal-state management.The Wu et al. judge penalizes filtering superseded facts in knowledge-update questions, while lifecycle is approximately neutral on remaining task types.Source: §D.2 · p. 35, Table 21 · p. 37

D.3 Scaling Validation

• Lifecycle gains widen from 8 to 40 personas while staleness remains stable, and FR-Graphiti’s strongest attack-vector gains occur on superseded preferences, expired logistics, multi-version facts, and selective forgetting.The reported FR-Graphiti gains are +21pp on AV1, +20pp on AV2, +22pp on AV4, and +10pp on AV7.Source: §D.3 · p. 36, §D.7 · p. 37, Table 22 · p. 37

D.4 Substrate Independence

• FR-Graphiti transfers the lifecycle policies without modification, achieving 8% staleness versus FR-Bank’s 15.5%, while enhanced FR-Bank retrieval compensates for lacking graph traversal.This supports lifecycle management as a policy contribution rather than an infrastructure artifact.Source: §D.4 · p. 36

D.6 Bootstrap Confidence Intervals and Pairwise Significance

• +12.2pp full-versus-baseline gain, with full 72.9%, no-routing 66.7%, uniform decay 66.5%, and baseline 60.7%; all ablation gaps are statistically significant.Paired bootstrap intervals are +8.3 to +16.1pp for full versus baseline, +2.9 to +9.7pp versus no-routing, and +3.5 to +9.3pp versus uniform decay, with McNemar p<0.001 for the full-versus-baseline comparison.Source: §D.6 · p. 37

D.8 Efficiency and Positional Staleness

• FR achieves 3× lower stale-token rate and 8× lower stale-fact-at-rank-1 rate than Mem0, while controlled AV2 cases attribute opposite outcomes to expiry filtering.In 14 same-question, same-judge AV2 cases, FR suppressed expired facts that Mem0 surfaced with staleness=1.0.Source: §D.8 · p. 37, §D.10 · p. 38

D.9 Deterministic Lifecycle Layer Latency

• 47 μs median deterministic lifecycle latency at k=20, with a 1.96 μs per-candidate slope and near-theoretical linear scaling to k=1000.The lifecycle layer is three orders of magnitude below the bounded LLM distillation call and negligible against cosine-similarity work over stored edges.Source: §D.9 · p. 38

E.9 Architectural Prerequisites for Lifecycle Management

• Competitors cannot generally gain lifecycle behavior through metadata alone because they lack combinations of state-preserving updates, lifecycle-aware retrieval hooks, and query-intent classification.Mem0 overwrites prior vector payloads, Memory-R1 has fixed storage and action contracts plus pure-cosine retrieval, and MemoryOS loses fact-level links during consolidation; adding lifecycle features would reproduce FR’s pipeline.Source: §E.9 · pp. 40–42

F Slot-Key Supersession: Extended Examples

• Per-category time decay would require modifying the trained policy’s reward signal.Source: F Slot-Key Supersession: Extended… · p. 43

G Multi-Reference-Frame Temporal Sensitivity

• Temporal sensitivity uses absolute time, session-relative time, and conversational frequency, with routing specialized by behavioral category.Logistical facts use calendar time, Relational and Hobby facts use session gaps, Obligations use deadlines, Identity is nearly insensitive, and Preferences use absolute time.Source: G Multi-Reference-Frame Temporal… · p. 43

H Numeric Preservation: Detailed Mechanisms

• AV8 recovery improved from 81% to 94% through exemption from expiry and supersession filters plus a 0.95× reranking floor for high-scoring numeric edges.The supplement identifies deduplication resolution and arbitrary smallest-UUID canonical selection as numeric-loss points.Source: H Numeric Preservation: Detailed… · p. 43

Reproducibility Supplement

• The appendices document development, parameter analysis, and per-mechanism findings, while code, evaluation results, and run logs are publicly archived at DOI 10.5281/zenodo.20067778.Source: Reproducibility Supplement · p. 43

I Rate Calibration: The Squatting Phenomenon

• Under the initial 800× rate configuration, rates spanned from λ=0.0001 for Identity to λ=0.080 for Logistical.Source: I Rate Calibration: The Squatting… · p. 43
• The ontology’s competitive mid-alpha performance arose because its slowest rate, Procedural at λ=0.003, prevented multi-month category squatting, matching explicit calibration.The rate floor prevents any category from permanently dominating ranking positions.Source: I Rate Calibration: The Squatting… · p. 43
• Ten of eleven categories tolerated 0.25×–4× decay-rate sweeps, while Relational_Bonds tolerated 0.33×–4.0×, with every default interior to its feasible region.Feasibility required mean Jaccard@10 ≥0.95 over 516 LifecycleBench questions, indicating no category sat on a sensitivity cliff.Source: I Rate Calibration: The Squatting… · p. 44, Table 36 · p. 44

J.2 Between-Phase Improvements

• Phase 2→3 improvements scaled evaluation from 8 to 20 personas and added per-source normalization, cross-category supersession, and a world-knowledge filter reducing contamination from 86% to 1.4%.These changes accompanied the transition to Phase 3.Source: §J.2 · p. 44
• Phase 3→4 added scaling to 40 personas, retraction extraction, AV8 numeric rescue, date-aware expiry filtering, and parallel evaluation for a 12× speedup.The improvements target both memory handling and evaluation efficiency.Source: §J.2 · p. 44
• Phase 4→5b added multi-stage numeric rescue, backward-looking expiry bypass, and a future-preference guard for date resolution.These changes further addressed numeric and temporal disambiguation failures.Source: §J.2 · p. 44

J.1 Phase 2: Initial Validation (8 personas, 112 questions)

• Phase 2 per-AV results are reported for the initial validation stage, while Phase 3 overall results are reported for 20 personas and 257 questions.The supplied material identifies these result tables but does not provide their cell values.Source: Table 37 · p. 45, Table 38 · p. 45

Backward-looking query detection

• Backward-looking query detection bypasses supersession filtering for phrases such as “before switching” and “used to,” recovering 3 otherwise unreachable questions.Source: K Detailed Findings from… · p. 45

Category-specific α

• Category-specific decay reduces AV8 performance to 12.5% under uniform α=0.3, whereas Financial and Logistical use α values of 0.05 and 0.40.Source: K Detailed Findings from… · p. 45

Semantic floors

• Hit@1 rises from 19% to 31% when semantic floors preserve near-perfect matches for stable categories, using floors of 0.97 for Financial and 0.00 for Logistical.Zero activation would otherwise bury semantically strong matches in stable categories.Source: K Detailed Findings from… · p. 45

Pass-by-absence

• Pass-by-absence affects both systems on AV2: FR records 50 cases through expiry filtering, while Mem0 records 29 through extraction failure.Source: K Detailed Findings from… · p. 45

AV3 tradeoff decomposition

• AV3 errors decompose into 77% extraction misses, 12% decay-killed cases, and 12% partial failures, making extraction the dominant limitation.Source: K Detailed Findings from… · p. 45

L Scalability Analysis

• 3.0% stale tokens for FR versus 9.0% for Mem0 saves $25.12 per million queries in stale-token costs alone.The savings use GPT-4o pricing.Source: Table 39 · p. 45, L Scalability Analysis · p. 46

M Composability

• FR’s composable policy layer supports incremental adoption, behavioral extensions, and composition with Graphiti and FluxMem without requiring data migration.FluxMem’s structure selection and FR’s lifecycle policy are described as orthogonal.Source: M Composability · p. 46
• The AV3 gap narrows from −16pp to −9pp when supplementary extraction improves, while the lifecycle layer remains unchanged.The result indicates that lifecycle-managed retrieval immediately benefits from better extraction.Source: M Composability · p. 46

N.1 Methodology

• The evaluation feeds each system’s top-10 facts for 516 questions to GPT-5.4 under a fact-restricted prompt, with Claude Sonnet judging correctness and confabulation.Memory pipelines use versioned gpt-4.1-mini, while Mem0 default extraction uses gpt-4.1-nano; LongMemEval-S uses the Wu et al. protocol with gpt-4o-2024-08-06.Source: §N.1 · p. 46

N.2 Staleness-Confabulation Causal Chain

• 72–78% confabulation occurs with stale retrieval contexts versus 18–30% with clean contexts, linking lifecycle filtering to downstream answer quality.Stale contexts contain any outdated fact in the top-10 retrieved facts.Source: §N.2 · p. 46, Table 40 · p. 46

N.3 Reasoning Token Overhead

• FR’s confabulation gap is a lifecycle-architecture effect rather than test-time compute: stale contexts increase reasoning tokens by only ≈10% on GPT-5.4.Table 41 reports reasoning-token overhead by context staleness.Source: §N.3 · p. 46, Table 41 · p. 46

N.4 Abstain Analysis

• FR-Bank abstains on 42.1% of queries versus Mem0’s 28.7%, but the extra abstentions are drawn almost entirely from would-be wrong or fabricated answers.FR-Graphiti abstains on 44.0%, adding 79 abstentions over Mem0, while FR-Bank adds 69.Source: §N.4 · pp. 46–47

N.7 Case Studies

• 14 cases show FR producing CORRECT responses where Mem0 produces WRONG responses with confabulation by filtering outdated facts before answer generation.Examples include retracted vaping, a superseded vehicle, and an abandoned business plan; the supplied passages provide no numeric aggregate for N.5 or N.6.Source: §N.7 · p. 47, Table 42 · p. 47, Table 43 · p. 47

Architectural mechanisms

• 93% versus 65% on AV2 and 40% versus 5% on AV7 show FR’s event-time expiry and retraction filtering outperform Mem0’s flat-store architecture.These capabilities require structural support rather than improved prompting.Source: §N.8 · p. 48

Learned operations

• Memory-R1 reaches 0.816 MRR and 94% on AV3 but trails FR-Bank on event-time expiry, retraction, and numeric preservation because its action vocabulary lacks those mechanisms.FR’s ontology supplies the structural prior, while reinforcement learning can optimize when to apply lifecycle policies.Source: §N.8 · p. 48

Scaling and limitations

• AV7 improved from 0% in Phase 2 to 40% in Phase 5c across five phases, while supplementary identity/health extraction narrowed the AV3 gap from −16pp to −9pp, confirming composability.Assistant-generated content remained weakest at 75.4% pass@10, with 78.6% maximum, up from 50.0% before the described changes.Source: §N.8 · pp. 48–49

Staleness and confabulation

• 72–78% of stale-context responses are contaminated versus 18–30% with clean context, indicating that FR primarily improves outcomes by controlling the retrieved input distribution.On AV7, Mem0 causes 91.4% confabulation compared with FR-Bank’s 33.3%, while Memory-R1 remains at 90.0%.Source: §N.8 · pp. 48–49

Generator robustness

• FR-Bank’s safe-response ranking remains highest across generators, but Kimi K2.5 is 6.9–12.3pp less safe than GPT-5.4 across all systems.FR-Bank declines from 73.3% to 61.0%, consistent with Kimi’s stronger answering preference without a commensurate correctness gain.Source: §N.8 · p. 49

Cross-benchmark validation

• 75.2% on LongMemEval-S, 76.9% on LifecycleBench, and 79.5% on the LongMemEval oracle subset establish competitive retrieval and lifecycle-aware performance without tradeoff.No existing system matches this combination across standard retrieval and temporal-disambiguation benchmarks.Source: §N.8 · p. 49

O Mem0 Comparison Methodology

• Mem0 used its v1.0.5 defaults with gpt-4.1-nano extraction, text-embedding-3-small embeddings, Qdrant, disk persistence, raw turns, and one Claude Sonnet judge across 40 personas and 1,400 sessions.Frequent JSON parse errors during UPDATE and DELETE caused some lifecycle updates to fail silently.Source: O Mem0 Comparison Methodology · p. 49
• 67.1% AV-pass versus 61.0% for default Mem0, alongside 33.8% versus 45.1% confabulation, shows that a stronger extractor helps but does not close architectural gaps.Expired logistics remained 61% versus FR-Bank’s 88.0%, and selective forgetting remained 20% versus FR-Graphiti’s 40%; stronger-extractor Mem0 matched Memory-R1 at 67.1% versus 66.9%.Source: O Mem0 Comparison Methodology · p. 49

P Memory-R1 Comparison Methodology

• Memory-R1 was reimplemented with GPT-4.1-mini agents as an upper bound on pre-RL performance, GPT-4o-mini extraction, 1536-dimensional embeddings, 60 retrieved candidates, and 10 distilled memories.Across 40 personas, it processed 65,411 operations and retained 4,278 entries, with NONE accounting for 89.3% of operations.Source: P Memory-R1 Comparison Methodology · p. 50
• MemoryOS uses hierarchical episodic, semantic, and procedural tiers but requires approximately 1,500 LLM calls per 35-session persona versus FR-Bank’s ∼50, a 30× ingestion compute differential.The comparison frames Memory-R1’s learned-policy objective and Fortunate Recall’s interpretable-policy objective as complementary.Source: P Memory-R1 Comparison Methodology · p. 50

P.1 A-MEM Comparison Methodology

• 65.3% pass, 87.6% Hit@5, 30.4% staleness, and 47.0% confabulation make A-MEM’s high recall insufficient without lifecycle management.A-MEM was evaluated with gpt-4.1-mini, the default evolution threshold, and text-embedding-3-small; its independently designed failure pattern supports benchmark-independence.Source: §P.1 · p. 50

Q Fortunate Recall–Specific Metrics

• FR’s proposed positional metrics report 14.0% versus Mem0’s 37.2% positional staleness exposure and 0.51 versus 1.84 weighted staleness risk.The weighted risk assigns attention weights of 3.0, 1.0, and 2.0 to ranks 1–3, 4–7, and 8–10, respectively.Source: Q Fortunate Recall–Specific Metrics · p. 51

R Literature Positioning; S Full Alpha Sweep; T Fine-Grained Alpha Sweep with Statistical Tests; U Category-Specific Parameters; V Decay Rates

• The evaluation reports literature positioning, a full and fine-grained α sweep with paired significance tests, category-specific blending weights and semantic floors, and per-category decay rates.The fine-grained sweep’s table caption states that all p-values fail Bonferroni correction at p< 0.0056; the decay-rate table reports a 5.3× spread.Source: Table 45 · p. 51, S Full Alpha Sweep · p. 51, Table 46 · p. 51

W Cross-System Outcome Matrix and Positional Staleness; X Ablation Summary; Y Positional Utilization at Retrieval Depth

• The cross-system outcome matrix, positional-staleness analysis, component ablation summary, and retrieval-depth analysis evaluate FR–Mem0 outcomes and positional effects, with no significant positional effect on 2026 frontier models.At 50-document scale on Claude Sonnet (2026), the replication found point-biserial r= −0.009, p= 0.65, 95% CI [−0.044, +0.022].Source: W Cross-System Outcome Matrix and… · p. 53, Table 50 · p. 53, Table 51 · p. 53

Z FR-Bank Architecture and Methodology

• FR-Bank’s standalone bank uses GPT-4.1-mini for combined extraction and classification, instance-qualified slot keys, category-dependent thresholds, and an eight-stage retrieval pipeline.The pipeline combines cosine, BM25, category-forced retrieval, deduplication, staleness-aware distillation, lifecycle filters, blended scoring, and top-10 output; the released bank contains 18,936 entries, with 31.5% deactivated.Source: Z FR-Bank Architecture and Methodology · p. 54

Kimi K2.5 Cross-Generator Validation

• The rank ordering FR-Bank < FR-Graphiti < Memory-R1 < Mem0 replicates on Kimi K2.5, with lifecycle-managed systems retaining an approximately 15pp tier-average advantage.The comparison is a stress test rather than a same-temperature replication because Kimi uses its default temperature of 1.0; Kimi is also approximately 10pp less safe overall.Source: Kimi K2.5 Cross-Generator Validation · p. 55

AB.1 Full Per-Attack-Vector Breakdown (Kimi K2.5)

• On Kimi K2.5, FR-Bank reaches 51.4% correct with 6.1% confabulation on AV1 and 19.7% correct on AV2, while its AV9 correctness falls to 10.0%.The AV9 limitation arises when Kimi commits between simultaneously retrieved values under soft supersession; AV2 performance is attributed to event-time expiry filtering.Source: AB.1 Full Per-Attack-Vector Breakdown… · p. 56

Extended Discussion

• LifecycleBench supports the thesis through precision: FR reaches 73% pass with 8% staleness at 40-persona scale, while the baseline reaches 61% pass with 18% staleness.The same policy layer transfers across Graphiti and the embedding bank, although FR-Bank trades higher recall for 15.5% staleness; staleness is most harmful when recall is low.Source: Extended Discussion · pp. 57–58
• 76.9% pass rate and 22.4% confabulation make FR-Bank the strongest system in the five-system comparison, ahead of Mem0, Memory-R1, and MemoryOS.MemoryOS reaches 70.5% pass but only 13.2% correct rate, illustrating the gap between retrieval metrics and downstream response quality.Source: Extended Discussion · p. 57

LongMemEval FR-Bank Evaluation Methodology

• 75.2% on full LongMemEval-S and 76.9% on LifecycleBench cover complementary regimes: oracle per-question retrieval and persistent per-persona lifecycle management.The LongMemEval-S result uses the exact Wu et al. rubric, while LifecycleBench evaluates 516 lifecycle-sensitive queries against one accumulating store with 15.5% staleness.Source: LongMemEval FR-Bank Evaluation… · p. 54, Extended Discussion · p. 58

AC.1 Privacy, Consent, and Safety

• Deployment requires consent, verifiable deletion, poisoning defenses, and stronger prompt-injection protection, while sensitive-category persistence and culturally variable decay assumptions remain unvalidated.The interpretable ontology provides a foundation for user-facing controls, but this work does not evaluate such interfaces or validate per-user parameter adaptation.Source: AC.1 Privacy, Consent, and Safety · pp. 58–59

AD.2 What the Arm Is, Exactly

• The untyped three-primitive arm retains supersession, expiry, and retraction masking but removes ontology, routing, activation, decay, and score blending, so it lower-bounds lifecycle metadata alone.In the paired 516-question evaluation, correctness is null while 84 questions are confabulated only by the untyped arm versus 21 only by the full stack, alongside a 15.9pp abstention increase.Source: AD.2 What the Arm Is, Exactly · p. 59, AD.3 Protocol and Results · p. 60

AE.4 Honest Negatives

• 76 vs. 75 confabulation counts make the minimal and full FR arms indistinguishable on re-execution, so the minimal-to-full confabulation drop is unreplicated.The Mem0-to-FR confabulation drop and the correctness component do replicate.Source: AE.4 Honest Negatives · p. 61

AE.3 Replication and Spread / AE.4 Honest Negatives

• 135/280 lifecycle-off questions beat both typed arms overall, locating the knowledge-update cost in generic lifecycle primitives rather than the ontology.Over-broad slot keys deactivated current values, emptying the candidate pool in 7 cases; FR-full also varied from 125–131/280 across three runs.Source: AE.4 Honest Negatives · p. 61, AE.3 Replication and Spread · p. 61

All-Queries End-to-End Metrics

• 31.2% vs. 18.6% correct-over-total favors FR-Bank over Mem0, and the ranking remains essentially stable after changing the denominator.The comparison uses all 516 queries; 19.2 of the 22.7pp FR-Bank–Mem0 gap survives the denominator change.Source: All-Queries End-to-End Metrics · p. 62

AG.1 Six-Model Cross-Family Panel

• 19 of 24 pinned partial verdicts moved elsewhere under the six-model panel, revealing that the pinned judge treats incompleteness as partial rather than contamination.Correct, wrong, and abstain verdicts were otherwise stable against the panel majority.Source: AG.1 Six-Model Cross-Family Panel · p. 62

AG.2 Targeted Re-Judge of Every PARTIAL Verdict

• +5.2 to +7.0pp corrected correct rates leave the cross-system ranking unchanged, while narrowing the FR-Bank–Mem0 gap from +12.6 to +11.0pp.The independent re-judge corrected all 748 partial verdicts across six systems; the two audits agreed on bias direction but differed in magnitude.Source: AG.2 Targeted Re-Judge of Every… · p. 63

AG.3 What Remains / AD.4 Judge-Pass Provenance

• Absolute rates remain judge-relative, while cross-system comparisons are robust; the planned stratified human-labeled audit has not yet been run.The pinned-judge results remain the paper’s numbers of record because every system was scored identically.Source: AD.4 Judge-Pass Provenance · p. 60, AG.3 What Remains · p. 63

Metadata-Noise Sensitivity Across the Full Basis

• Kendall τ≈0.997 while Jaccard falls to 0.76 under slot-key corruption, showing that metadata noise changes set composition more than ranking.Category labels remain comparatively tolerant, whereas retraction states are potent per edge but rare; event-time corruption is mostly inert near current expiry boundaries.Source: Metadata-Noise Sensitivity Across the… · p. 63, Table 58 · p. 64
• 37.5% end-to-end AV7 correctness is FR-Bank’s highest evaluated-system result, despite only 2/40 retractions being cleanly suppressed at the metadata level.Selective forgetting is limited by both retraction-detection recall and slot-key fragmentation, with 38/40 events leaking at least one live edge.Source: Metadata-Noise Sensitivity Across the… · p. 64

NeurIPS Paper Checklist

• The study involved no human subjects: personas were synthetic, and classifier-agreement audits used LLM judges rather than human annotators.Accordingly, participant instructions, compensation, risks, and IRB approvals were marked not applicable.Source: NeurIPS Paper Checklist · pp. 67–68
• The Fortunate Recall pipeline uses gpt-4.1-mini for per-turn entity, edge, and behavioral classification, retrieval distillation, and documents its boundary with lifecycle mathematics.Answer generation uses GPT-5.4 primarily and Kimi K2.5 for cross-generator validation; the boundary is documented in §3 and Figure 1.Source: NeurIPS Paper Checklist · p. 68