Search papers, labs, and topics across Lattice.
×\times compression, the most aggressive loss observed. The variation in degradation curves across models reflects differences in summarization strategy, not differences in vulnerability: all models converge to near-zero accuracy at high compression ratios. KO retrieval achieves 100% across all models at all ratios, since it bypasses compaction entirely. This cross-model replication establishes compaction loss as an architectural property of prose-based memory, not a limitation of any particular model. Lossy compression of fact-dense content is fundamentally unable to preserve arbitrary precision regardless of the model performing the compression. 7 Knowledge Objects 7.1 Architecture Knowledge Objects are not a replacement for LLMs—they are an external memory layer that LLMs read from and write to. The LLM remains the reasoning engine; KOs provide the storage substrate. A Knowledge Object is a discrete, hash-addressed tuple: KO=(subject,predicate,object,provenance_metadata)\text{KO}=(subject,predicate,object,provenance\_metadata) (4) The provenance metadata enables downstream attribution and audit; in systems where contributors should be compensated for their knowledge, this metadata can feed into attribution-native learning pipelines (Zahn et al., 2026b). The hash key is computed deterministically: key=SHA-256(normalize(subject)∥normalize(predicate))key=\text{SHA-256}(\text{normalize}(subject)\|\text{normalize}(predicate)) (5) This gives O(1)O(1) lookup: a query is parsed into (subject,predicate)(subject,predicate), hashed, and the corresponding KO is retrieved from a simple key-value store (SQLite or PostgreSQL). Storage cost is ∼\sim100 bytes per KO, negligible at any scale. 7.2 Query Pipeline The KO query pipeline consists of two LLM calls: 1. Parse (Haiku, ∼\sim500 input tokens): extract (subject,predicate)(subject,predicate) from the natural language query. 2. Answer (Sonnet, ∼\sim300 input tokens): generate an answer given the retrieved KO. Total tokens per query: ∼\sim900, regardless of corpus size. This is the source of the O(1)O(1) cost property. 7.3 Robustness to Noisy Contexts A natural concern with the KO pipeline is whether the parsing step can reliably extract (subject,predicate)(subject,predicate) tuples from realistic, unstructured inputs rather than clean benchmark queries. To address this, we tested parsing accuracy across five input conditions: Table 6: KO parsing robustness across input conditions 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% Context noise does not affect KO parsing: when facts are embedded within dense clinical abstracts, conversational messages, or pronoun-heavy text requiring coreference resolution, the Haiku parser successfully extracts the correct (subject,predicate)(subject,predicate) tuple in all cases tested. This robustness stems from the LLM’s strong natural language understanding—the parser does not rely on surface patterns but on semantic comprehension of what is being asked. However, query phrasing variations present a genuine vulnerability. When users express the same query using non-standard predicate formulations (e.g., “Ki value of Erlotinib against EGFR” instead of “Ki for EGFR inhibition”), the parser may return a predicate format that does not match the stored key, causing a lookup miss (20% failure rate on adversarial phrasing). This is not a fundamental architectural weakness but a predicate normalization issue addressable through synonym mapping or fuzzy matching on the predicate field—engineering improvements we leave for future work. 7.4 Ingestion KOs can be ingested from structured data (schema mapping, trivial) or extracted from conversations in real time using an LLM. When a user says “we decided against Redis,” the LLM can simultaneously participate in the conversation and emit: KO: (caching_technology, vetoed, Redis) provenance: "team meeting, 2025-01-15" Cost: $0.36 per 1,000 facts using Haiku, recouped in ∼\sim5 queries (at N=1
1
0
3
LLMs forget up to 60% of facts when summarizing and erode over half of project constraints during iterative compaction, but a simple discrete memory system (KOs) fixes this while slashing costs by 252x.
