Every AI system that learns from interactions faces the same fundamental tension: storing everything is expensive, but forgetting too much destroys value.
Most systems resolve this by deciding what to keep and what to discard — losing information in the process. The Delta Engine takes a different approach: it stores everything, but in a radically more efficient form.
Instead of storing the full state of every interaction, the Delta Engine stores only the difference between states — the delta. Full state can always be reconstructed from any checkpoint by applying all subsequent deltas. But normal read operations access only the delta stream, not the full state, making recall dramatically faster and cheaper.
How Delta Storage Works
Traditional AI memory stores full conversation snapshots:
State 1: { query: "What is FDIA?", context: [], response: "..." }
State 2: { query: "Explain D variable", context: [State1], response: "..." }
State 3: { query: "How does A gate work?", context: [State1, State2], response: "..." }
Storage grows as O(n²) — each state contains all previous context.
Delta Engine stores only the changes:
Checkpoint: { query: "What is FDIA?", response: "..." }
Delta 1: { added_context: [Checkpoint], query: "Explain D variable", response: "..." }
Delta 2: { added_context: [Δ1], query: "How does A gate work?", response: "..." }
Storage grows as O(n) — each delta only contains the incremental change.
Result: 74% compression vs full-state storage, measured across the RCT Ecosystem production workload (v5.4.5).
The 7-State Recall Pipeline
The Delta Engine is tightly integrated with the Intent Loop Engine's 7-state pipeline:
Cold Start Path (3–5 seconds)
RECEIVED → VALIDATED (FDIA) → MEMORY_CHECK → [MISS] → COMPUTING → VERIFYING → COMMITTING → COMPLETED
- Query arrives
- FDIA validates quality and intent
- Delta Engine checks RCTDB for a warm cache hit (semantic similarity > 0.95)
- Cache miss: full computation path activated (3–5s)
- HexaCore selects optimal model(s)
- SignedAI verifies output
- Result committed to Delta Engine store with full provenance
Warm Recall Path (<50ms)
RECEIVED → VALIDATED (FDIA) → MEMORY_CHECK → [HIT] → COMPLETED
- Query arrives
- FDIA validates quality and intent
- Delta Engine checks RCTDB — semantic similarity > 0.95 found
- Cache hit: cached response returned in <50ms
- Cost: near-zero (no LLM call)
As the system is used, more queries hit the warm path. Over time, cost per query trends toward zero for stable, repeating patterns.
Semantic Similarity: How the Cache Finds Matches
The Delta Engine does not cache exact strings. It caches semantic embeddings — mathematical representations of meaning.
Two queries can be semantically equivalent even if the wording differs completely:
| Query 1 | Query 2 | Similarity | |---|---|---| | "What is FDIA?" | "Explain the FDIA equation" | 0.97 (same intent) | | "How does constitutional AI work?" | "What is constitutional AI?" | 0.95 (similar intent) | | "FDIA vs RLHF" | "JITNA protocol explained" | 0.21 (different topics) |
The similarity threshold of 0.95 is configurable per use case. Higher thresholds (stricter matching) reduce false positives but lower cache hit rates. Lower thresholds (looser matching) improve hit rates but risk returning slightly off-target responses.
8-Dimensional RCTDB Schema
Delta Engine storage uses RCTDB — an 8-dimensional universal memory schema. Each record in RCTDB has 8 fields that enable recall, audit, and deletion:
| Dimension | Description | |---|---| | query_hash | SHA-256 hash of semantic embedding — enables O(1) lookup | | fdia_scores | D, I, A values at time of storage | | subject_uuid | Links record to data subject for PDPA erasure | | model_chain | Which model(s) produced the output | | consensus_result | SignedAI tier and voting outcome | | delta_chain | Reference to parent checkpoint/delta | | timestamp | ISO 8601 creation time | | provenance | Data sources, lawful bases, jurisdiction zones |
The 8-dimensional schema enables queries that cross all 8 dimensions simultaneously — faster than naive joins across separate tables.
Three Storage Zones
RCTDB operates across 3 zones with different performance characteristics:
| Zone | Technology | Latency | Use Case | |---|---|---|---| | Hot | In-memory (Redis-compatible) | <1ms | Recent queries, active sessions | | Warm | Fast SSD | 1–5ms | Past 30 days of queries | | Cold | Compressed blob | >10ms | Archive, audit, long-term compliance |
Queries automatically migrate between zones based on access frequency. A frequently accessed query moves from Cold to Warm to Hot. An infrequently accessed query moves from Hot through Warm to Cold.
Integration with FDIA
The Delta Engine is connected to the FDIA equation at two points:
At MEMORY_CHECK (State 3 of 7-State Pipeline): The Data quality score (D) in FDIA partially reflects the cache freshness of relevant information. A warm cache hit with high semantic similarity contributes positively to D — the system has recent, high-quality information about this specific topic.
At COMMITTING (State 6 of 7-State Pipeline): Before storing a new delta, the system validates the F score of the response being stored. Only responses that meet the minimum F threshold are stored in the warm/hot zones. Responses below threshold are stored in cold storage only — preventing low-quality information from contaminating the warm cache.
Performance Summary
| Metric | Value | |---|---| | Memory compression | 74% lossless | | Cold start latency | 3–5 seconds | | Warm recall latency | <50 milliseconds | | Semantic similarity threshold | 0.95 (configurable) | | Storage zones | 3 (Hot/Warm/Cold) | | RCTDB dimensions | 8 | | Recall accuracy | 96.1% (matches vs expected) |
Frequently Asked Questions
Is Delta Engine compression lossy?
No. Delta storage is lossless — full state can always be reconstructed from any checkpoint by applying all deltas in sequence. No information is lost; it is simply stored more efficiently.
How does Delta Engine compare to RAG (Retrieval-Augmented Generation)?
RAG retrieves information from a document store and injects it into the LLM context. Delta Engine caches complete, verified responses and returns them directly. For repeated queries, Delta Engine is faster and cheaper than RAG (no re-retrieval, no re-generation). For novel queries, both systems compute from scratch.
Can Delta Engine handle Thai language?
Yes. The semantic similarity engine uses multilingual embeddings that cover Thai, English, and all 8 language pairs supported by the RCT Ecosystem. Thai queries and Thai-language cached responses are fully supported.
Summary
The Delta Engine makes the RCT Ecosystem's performance economics sustainable:
- 74% compression through delta-only storage (O(n) vs O(n²))
- <50ms warm recall through semantic similarity matching (threshold: 0.95)
- Near-zero cost for stable, repeating query patterns over time
- RCTDB 8-dimensional schema for structured storage, audit, and PDPA-compliant deletion
- 3-zone architecture for intelligent hot/warm/cold data placement
This article was written by Ittirit Saengow, founder and sole developer of RCT Labs.
สิ่งที่องค์กรควรสรุปจากบทความนี้
The Delta Engine is the memory compression and recall system at the core of the RCT Ecosystem. By storing only state changes (deltas) rather than full state snapshots, it achieves 74% lossless compression and enables warm recall in under 50 milliseconds — reducing per-request cost to near zero for repeated patterns.
เชื่อมจากความรู้ไปสู่การประเมินระบบจริง
ทุกบทความเชิงวิจัยควรเชื่อมต่อไปยัง solution page, authority page, และ conversion path เพื่อให้การอ่านไม่จบแค่ traffic
บทความก่อนหน้า
Constitutional AI vs RAG: Which Architecture Actually Prevents Hallucination?
RAG (Retrieval-Augmented Generation) reduces hallucination by grounding responses in retrieved documents. Constitutional AI prevents hallucination through architectural constraints. This comparison explains the fundamental difference, performance data, and when to use each approach — or both.
บทความถัดไป
Evaluation Harnesses for Enterprise LLMs: Beyond Vibe-Testing
Most AI teams evaluate their LLM deployments by looking at outputs and deciding if they seem right. This is vibe-testing. Here is a rigorous alternative — how the RCT Ecosystem runs 4,849 automated tests across 8 evaluation levels to produce verifiable enterprise trust signals.
Ittirit Saengow
Primary authorอิทธิฤทธิ์ แซ่โง้ว คือผู้ก่อตั้ง นักพัฒนาเพียงคนเดียว และผู้เขียนหลักของ RCT Labs — แพลตฟอร์มระบบปฏิบัติการ AI แบบ constitutional ที่สร้างขึ้นอย่างอิสระตั้งแต่สถาปัตยกรรมจนถึงการเผยแพร่ เขาคิดค้นสมการ FDIA (F = (D^I) × A) ข้อกำหนดโปรโตคอล JITNA (RFC-001) สถาปัตยกรรม 10 ชั้น ระบบ 7-Genome และกระบวนการ RCT-7 แพลตฟอร์มทั้งหมด ทั้งโครงสร้างสองภาษา ระบบ SEO ระดับองค์กร ไมโครเซอร์วิส 62 ตัว อัลกอริทึม 41 ชุด และงานวิจัยทั้งหมดที่เผยแพร่ สร้างโดยคนเพียงคนเดียวในกรุงเทพฯ ประเทศไทย