Retrieval-augmented generation stores all content indiscriminately, degrading accuracy as noise accumulates. Parametric approaches compress knowledge into weights, precluding selective updates. Neither mirrors biological memory, which gates encoding based on salience and archives rather than deletes superseded information. We introduce write-time gating that filters incoming knowledge objects using composite salience scores (source reputation, novelty, reliability) while maintaining version chains that preserve prior states. Using real LLM evaluation without oracle access to quality labels, write gating achieves 100 percent accuracy versus 13 percent for ungated stores. The critical finding emerges under distractor scaling: at 8:1 distractor ratios, read-time filtering (Self-RAG) collapses to 0 percent while write gating maintains 100 percent, revealing a structural advantage of write-time over read-time curation. Validation on Wikipedia (20 entities), procedurally generated pharmacology data, and 2026 arXiv papers confirms these findings. The gating advantage scales inversely with parametric memory support: +25pp for Wikipedia, +48pp for post-cutoff arXiv, +65pp for procedural data with zero training knowledge. Signal ablation confirms the method does not depend on oracle-correlated metadata. Write gating matches Self-RAG accuracy at one-ninth the query-time cost.
Figures & tables
Figure 1: Complementary Learning Systems in biology and AI. The brain uses fast hippocampal storage for specific episodes and slow neocortical storage for semantic patterns. Our architecture mirrors this with discrete Knowledge Objects (fast, addressable) complementing LLM weights (slow, distributed).
System
Discrete
Write Gate
Archive
Versions
Provenance
RAG
✓
Titans
✓
DNC
✓
✓
MemGPT
✓
MaRS
✓
✓
✓
NEC
✓
Table 1: Capability comparison across memory architectures, based on published system descriptions. Our approach (SGW-HA) uniquely combines discrete objects with write-time gating, archiving, and version chains.
Figure 2: Write-time salience gating architecture. Incoming knowledge objects are scored on three observable signals (reputation, novelty, source reliability) without oracle access. Objects above threshold τ enter the active store; objects below are archived in cold storage. On our benchmark: 50 KOs → 13 admitted → 100% accuracy (vs 13.3% ungated).
Condition
Store Size
Compression
Accuracy
Ungated
50±0
0%
13.3%±4.7%
Gated
13±1
75%
100%±0%
Table 2: Write gating performance (mean ± std over 3 seeds, real LLM evaluation). Reputation-based gating achieves perfect accuracy without oracle access, storing only 25% of candidates.
Figure 3: Version chains preserve temporal history through supersession links rather than overwrites. When facts update, prior versions are archived with pointers, enabling temporal queries that overwrite-based systems cannot answer.
Method
Accuracy
No filtering
12.5%±0.0%
Self-RAG (read-time)
93.8%±6.3%
Write gating (write-time)
100%±0.0%
Both (write + read)
93.8%±6.3%
Table 3: Write-time vs. read-time filtering (mean ± std, real LLM evaluation). Write gating outperforms Self-RAG by +6.2 pp. Adding Self-RAG to a gated store slightly degrades performance.
Distractor Ratio
Ungated
Self-RAG
Write Gating
2:1
100%
100%
100%
4:1
16.7%
100%
100%
6:1
0%
100%
100%
8:1
0%
0%
100%
Table 4: Distractor ratio scaling (real LLM evaluation). Ungated retrieval collapses by 4:1. Self-RAG holds through 6:1 but collapses catastrophically at 8:1. Write gating maintains 100% at all tested ratios.
Figure 4: Accuracy versus distractor ratio. Ungated retrieval collapses monotonically. Self-RAG maintains a plateau through 6:1 before catastrophic collapse at 8:1. Write gating remains constant at 100% across all tested ratios. The shaded region indicates noise levels plausible for web-scale retrieval.
Figure 5: Four-way method comparison at 4:1 distractor ratio. Write gating outperforms Self-RAG by +6.2 pp. Combining both methods degrades to Self-RAG’s accuracy—the critic introduces false negatives in an already-clean store.
Condition
Accuracy
Store Size
Distractors Admitted
Ungated
85.2%±2.0%
365
292
Write-gated
98.1%±1.1%
102–110
29–37
Table 5: Write gating on Wikipedia facts (mean ± std over 5 seeds, 73 questions each, 4:1 distractor ratio). Write gating admits all 73 correct facts while filtering 87% of distractors.
Method
Accuracy
No filtering
84.9%±1.9%
Self-RAG (read-time)
97.7%±0.6%
Write gating (write-time)
96.8%±1.3%
Both (write + read)
98.2%±0.6%
Table 6: Write-time vs. read-time filtering on Wikipedia data (mean ± std, 3 seeds, 4:1 ratio). Both filtering approaches achieve > 96%; combining both reaches 98.2%.
Ratio
Ungated
Write Gating
Self-RAG
2:1
98.6%
98.2%
98.6%
4:1
84.9%
97.7%
97.3%
8:1
72.6%
97.7%
95.9%
Table 7: Wikipedia distractor ratio scaling (mean over 3 seeds). Ungated accuracy degrades from 98.6% to 72.6% as noise increases. Write gating and Self-RAG both maintain > 95% at all ratios.
Configuration
Accuracy
Store Size
Correct / Incorrect
Ungated
85.2%±2.0%
365
73 / 292
Full gating (rep + nov + label)
97.8%±1.1%
∼110
73 / 37
No label (rep + novelty only)
96.4%±1.4%
∼155
72 / 83
Table 8: Source label ablation on Wikipedia data (mean ± std over 5 seeds, 73 questions each, 4:1 distractor ratio). Removing the source label signal costs only 1.4pp.
Ratio
Ungated
Gated
Self-RAG
2:1
98.6%±1.4%
98.8%±1.2%
97.0%±1.7%
4:1
63.0%±5.2%
98.2%±1.0%
95.0%±2.4%
8:1
32.0%±4.0%
96.6%±1.6%
80.2%±2.1%
Table 9: Retrieval accuracy on procedural pharmacology data (5-seed means ± std, 100 questions per seed). Facts the LLM provably cannot know from training.
Ratio
Ungated
Gated
Self-RAG
2:1
94.8%±2.7%
98.0%±2.5%
98.8%±1.6%
4:1
71.6%±4.1%
97.6%±2.3%
94.4%±3.4%
8:1
45.2%±2.0%
93.6%±3.2%
83.2%±1.6%
Table 10: Retrieval accuracy on 2026 arXiv papers (5-seed means ± std). Knowledge provably absent from the model’s training data.
Step
Store Size
Gated Size
Ungated
Gated
1
250
∼50
68.0%
100.0%
5
1,250
∼248
62.0%
97.0%
10
2,500
∼496
63.0%
97.0%
Table 11: Retrieval accuracy as the store grows (5-seed means). Gated store admits ∼19.8% of candidates, maintaining 99.1% precision throughout.
Figure 6: Distractor scaling on novel-domain data (5-seed means ± std). (a) Pharmacology corpus with zero parametric memory support: ungated accuracy drops to 32% at 8:1 while write gating maintains 96.6% . (b) 2026 arXiv papers post-training-cutoff: ungated drops to 45.2% at 8:1 while write gating maintains 93.6% . In both domains, write gating degrades less than Self-RAG as noise increases.
Figure 7: Gating advantage at 8:1 distractor ratio across knowledge regimes. The advantage scales inversely with parametric memory support: +25.1 pp for well-known facts (Wikipedia), +48.4 pp for post-cutoff research (arXiv 2026), +64.6 pp for procedurally generated data with zero training knowledge (pharmacology). Gated accuracy remains 93 – 97% regardless of domain; the growing green region reflects increasing ungated degradation.
Configuration
Accuracy
Store Size
Full model (all 3 signals)
100.0% ± 0.0%
12.8
Source label only
100.0% ± 0.0%
10.0
Reputation + label
98.0% ± 4.0%
11.4
Reputation + novelty
98.0% ± 4.0%
16.4
Reputation only
94.0% ± 8.0%
12.6
Novelty only
58.0% ± 19.4%
15.8
Table 12: Signal ablation: accuracy by gating configuration (5-seed mean ± std)
Figure 8: Pattern separation in the dentate gyrus. Sparse coding (2–4% activation) forces similar inputs to have dissimilar representations, preventing interference. Mossy fiber connections are deprioritized rather than deleted, preserving recovery capability—the biological basis for archive semantics.
Corruption Rate
Gated
Ungated
0% (baseline)
98.0% ± 4.5%
4.0% ± 5.5%
5%
78.0% ± 16.4%
4.0% ± 5.5%
10%
74.0% ± 12.0%
2.0% ± 4.5%
20%
52.0% ± 20.5%
6.0% ± 5.5%
30%
30.0% ± 14.1%
0.0% ± 0.0%
Table 13: Corruption robustness: accuracy when verified sources contain incorrect facts (5-seed mean ± std, n=10 entities per seed).
Memory-augmented language models combine persistent or temporary memory with neural computation to retain, retrieve, and use information beyond the active context.