Facts as First Class Objects: Knowledge Objects for Persistent LLM Memory

Oliver Zahn; Simran Chana

cs.AIMar 18, 2026

Facts as First Class Objects: Knowledge Objects for Persistent LLM Memory

Authors: Oliver Zahn, Simran Chana

Organizations: Independent Researcher · University of Cambridge

Abstract

Large language models increasingly serve as persistent knowledge workers, with in-context memory - facts stored in the prompt - as the default strategy. We benchmark in-context memory against Knowledge Objects (KOs), discrete hash-addressed tuples with O(1) retrieval. Within the context window, Claude Sonnet 4.5 achieves 100% exact-match accuracy from 10 to 7,000 facts (97.5% of its 200K window). However, production deployment reveals three failure modes: capacity limits (prompts overflow at 8,000 facts), compaction loss (summarization destroys 60% of facts), and goal drift (cascading compaction erodes 54% of project constraints while the model continues with full confidence). KOs achieve 100% accuracy across all conditions at 252x lower cost. On multi-hop reasoning, KOs reach 78.9% versus 31.6% for in-context. Cross-model replication across four frontier models confirms compaction loss is architectural, not model-specific. We additionally show that embedding retrieval fails on adversarial facts (20% precision at 1) and that neural memory (Titans) stores facts but fails to retrieve them on demand. We introduce density-adaptive retrieval as a switching mechanism and release the benchmark suite.

Figures & tables

Method	Adversarial P@1	BEIR NDCG@10	Spurious Trigger Rate
Embedding only	$20.0\%$	0.645	—
Hybrid (always)	$100.0\%\pm 0.0\%$	0.619	100%
Adaptive ( $\tau{=}0.85$ )	$\mathbf{100.0\%\pm 0.0\%}$	0.645	0%

Table 1: Density-adaptive retrieval results (5 random seeds, mean

\pm

std)

N	Tokens	Claude IC	GPT-4o IC	RAG	KO
10	$\sim$ 0.3K	100%	— †	100%	100%
50	$\sim$ 1K	100%	— †	100%	100%
100	$\sim$ 3K	100%	— †	100%	100%
250	$\sim$ 7K	100%	— †	100%	100%
500	$\sim$ 14K	100%	— †	100%	100%
1,000	$\sim$ 27K	100%	19% ∗	100%	100%

Table 2: Scaling results: exact-match accuracy at each corpus size (30 queries per

N

for

N\geq 1{,}000

; 10–20 for smaller

N

). Claude IC and KO results replicated across 5 seeds with identical results except seed 123 (97% at

N{=}3{,}000

and

N{=}5{,}000

). RAG uses all-MiniLM-L6-v2 embeddings with top-5 retrieval.

Figure 1: Scaling curve: exact-match accuracy vs. corpus size ( $N{=}10$ to $N{=}10{,}000$ ). Claude Sonnet 4.5 maintains 100% accuracy through $N{=}7{,}000$ (97.5% of context window), then overflows. GPT-4o drops to 0% by $N{=}3{,}000$ . KO maintains 100% at all $N$ . The dashed line marks the 200K token context window boundary.

System	Correct	Abstained (lost)	Wrong
After compaction	40%	60%	0%
KO store	100%	0%	0%

Table 3: Fact retrieval after 36.7

\times

compaction (

N{=}2{,}000

)

Figure 2: Fact retrieval accuracy after 36.7 $\times$ compaction. In-context memory loses 60% of facts; KO maintains 100%.

Condition	Compression	Correct	Partial	Lost
1 $\times$ compaction	9 $\times$	91%	9%	0%
2 $\times$ compaction	17 $\times$	62%	1%	37%
3 $\times$ compaction	31 $\times$	46%	1%	53%
Full context	1 $\times$	100%	0%	0%
KO store	—	100%	0%	0%

Table 4: Goal preservation under cascading compaction (20 constraints, 5-seed mean)

Figure 3: Goal drift under cascading compaction. Left: Stacked bars showing correct, partial, and lost constraints after each round. Right: Accuracy decay vs. compression ratio. KO maintains 100% regardless of compaction.

Ratio	Sonnet 4.5	Sonnet 4.6	Opus	GPT-5.4
1 $\times$ (no compaction)	100%	100%	95%	100%
2 $\times$	25%	10%	80%	0%
5 $\times$	0%	20%	45%	20%
10 $\times$	0%	5%	15%	0%
20 $\times$	0%	5%	5%	5%

Table 5: Compaction accuracy across frontier models at varying compression ratios. KO achieves 100% across all models and ratios (omitted for space).

Condition	Parse Accuracy	End-to-End Accuracy
Clean query	100%	100%
Clinical abstract (PubMed-style paragraphs)	100%	100%
Conversational (Slack-style messages)	100%	100%
Coreference (pronoun-obscured subjects)	100%	100%
Messy query phrasing	80%	80%

Table 6: KO parsing robustness across input conditions

Experiment	In-Context	RAG	KO
Scaling ( $N{=}500$ )	1.00	1.00	1.00
Scaling ( $N{=}7{,}000$ )	1.00	1.00	1.00
Scaling ( $N{=}10{,}000$ )	overflow	1.00	1.00
Adversarial P@1	N/A	0.20	1.00
Fact compaction (37 $\times$ )	0.40	N/A	1.00
Goal drift (9 $\times$ )	0.91	N/A	1.00

Table 7: Summary: accuracy across all experimental conditions. N/A indicates the failure mode does not apply to that architecture.

Method	Accuracy
Model alone (no facts)	0.0%
Full context (all 500 facts)	31.6%
KO-grounded	78.9%

Table 8: Multi-hop reasoning accuracy (2-hop queries, 500-fact corpus,

n{=}19

)

Figure 4: Multi-hop reasoning accuracy on 2-hop queries over a 500-fact corpus. KO-grounded retrieval achieves 78.9% accuracy, a 47.3 percentage point improvement over full in-context presentation (31.6%).

Condition	Specificity	Groundedness	Actionability	Composite
Model alone (no KOs)	3.40	2.20	3.40	3.00
With KO retrieval	4.87	4.80	4.13	4.60

Table 9: Cross-domain synthesis quality (15 questions, 90 KOs across 3 domains)

Figure 5: Cross-domain synthesis quality scores (1–5 scale) across four dimensions. The largest improvement is in groundedness (+118%), where KO retrieval enables claims traceable to specific stored facts.

∗ Requires Opus at 5 $\times$ Sonnet pricing; Sonnet overflows at $N{>}7{,}400$ .
N facts	In-Context	KO	Ratio	Annual IC / KO
100	$0.009	$0.002	4 $\times$	$229 /$ 56
1,000	$0.082	$0.002	36 $\times$	$2,051 /$ 56
5,000	$0.406	$0.002	180 $\times$	$10,151 /$ 56
7,000	$0.568	$0.002	252 $\times$	$14,201 /$ 56
10,000	overflow ∗	$0.002	1,802 $\times$ ∗	$101,381 ∗ /$ 56

Table 10: Per-query cost and annual cost comparison (25,000 queries/year)

Figure 6: Production economics. Left: Per-query cost (log-log). In-context cost grows linearly with $N$ ; KO cost is flat. Vertical dashed lines mark context window limits. Right: Annual cost for a researcher (25,000 queries/year).

System	Capacity	Compaction	Goal Drift	Adversarial	Cost
In-context	fail ( $\sim$ 7K facts)	60% loss	54% loss	N/A	$O(N)$
RAG	OK	N/A	N/A	20% P@1	$O(k)$
KO	OK	OK	OK	100% P@1	$O(1)$

Table 11: Failure mode taxonomy across memory architectures

Evaluation	Titans	KO
Memorization (can facts be stored?)	100%	100%
Completion (free-form retrieval)	0–40%	100%
Forced choice (multiple choice)	45–85%	100%

Table 12: Titans neural memory vs. KO (dim=128, averaged over

N{=}10{,}25{,}50

)

Figure 7: Titans neural memory vs. KO across corpus sizes. While Titans achieves near-perfect memorization (light blue), retrieval accuracy degrades substantially on both forced-choice (red) and free-form completion (orange) tasks. KO maintains 100% accuracy regardless of corpus size.

Persistent Memory for Language Models

Persistent memory for language models studies how models retain, update, and use information across interactions and beyond their context windows.

See all 271 papers

Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs
Sep 9, 2026Ansuman Mullick, Eray TüzünConversational Memory Temporal Reasoning in Language Models
Bilkent University
Memory vs. Context? Influential Factors of Factual Recall in Language Models
Sep 21, 2026Guilhem Fouilhé, Nicholas Asher, Philippe MullerPrompt Sensitivity Memorization in Language Models
IRIT ANITI · IRIT CNRS · IRIT Université de Toulouse
EngramaBench: Evaluating Long-Term Conversational Memory with Structured Graph Retrieval
Apr 23, 2026Julian AcunaLLM Evaluation Memorization in Language Models

Sep 9, 2026J/K move · Enter open · S save

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

Ansuman Mullick, Eray Tüzün

Bilkent University

Conversational Memory Temporal Reasoning in Language Models Hallucination in Language Models Persistent Memory for Language Models Memory Agent Benchmarks Long-Term Conversational Memory

Open paper AskPDF arXiv

Current LLM memory systems treat all personal facts identically, so stores grow without bound while retrieval precision degrades. The core challenge is lifecycle management: which memories should persist, which should be replaced, and at what rate, conditioned on the behavioral type of each fact. Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle policies (differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing) as deterministic functions over LLM-extracted metadata. FR-Bank, our infrastructure-independent implementation, reaches a 76.9% pass rate on LifecycleBench, a new 516-question temporal-disambiguation benchmark, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%), and 75.2% on the full LongMemEval-S under the canonical Wu et al. judge protocol, so lifecycle policies impose no measurable cost on standard retrieval. A pre-registered ablation locates the gains: replacing the typed layer with three generic lifecycle primitives leaves correctness statistically unchanged (-1.7pp, 95% CI [-6.0, +2.7]), so the generic lifecycle metadata carries the correctness advantage, while the behavioral ontology carries calibration, halving downstream confabulation (12.0% vs 24.2%, p<0.001). End-to-end, FR-Bank cuts confabulation from Mem0's 45.1% to 22.4% over answered queries and from 32.2% to 13.0% over all queries while answering more of them correctly (31.2% vs 18.6%); the ranking replicates on the open-weight Kimi K2.5. The decomposition transfers to BEAM, an independently built benchmark: 46.8% correct vs Mem0's 32.9% over 280 questions, with the ontology's benefit concentrated in contradiction resolution and saturating near seven policy clusters. The ontology, benchmark, and code are released.

Abstract

Persistent Memory for Language Models

Explore similar work

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

Memory vs. Context? Influential Factors of Factual Recall in Language Models

EngramaBench: Evaluating Long-Term Conversational Memory with Structured Graph Retrieval