back to projects
case study: synapse
project log · recorded may '26
Synapse
#typescript#modelcontextprotocol(mcp)#sqlite#sqlite-vec#ripgrep#node.js
the vision
The minimalist transmission layer for AI context. Bio-inspired persistent memory, knowledge graph, and semantic recall designed to empower AI agents.
the hurdle
“AI agents frequently suffer from "context drift" in long-running projects. Standard RAG approaches often fail because they lack structural understanding (how components relate) and temporal context (why a decision was made). In complex codebases, naive line-based chunking cuts logical units in half, making the retrieved context "noisy" or incomplete for the AI agent.”
01the architecture
Engineered a "Local-First Project Cortex" using a decoupled layered architecture.
1. **Hybrid Retrieval Engine:** Combined BM25 (FTS5) for precise symbol matching with Vector Similarity Search (KNN) via `sqlite-vec` to handle semantic intent.
2. **AST-Aware Semantic Chunking:** Leveraged tree-sitter to parse code into atomic units (functions/classes), ensuring logical boundaries are preserved and enriched with parent-scope metadata.
3. **Temporal Knowledge Graph:** Designed a triple-store `(Subject, Predicate, Object)` system using Recursive CTEs for multi-hop relationship discovery and timestamped facts for "time-travel" debugging of architectural evolution.
02battle scars
**Challenge:** Balancing high-speed semantic recall with a "Zero Cloud" privacy mandate.
**Solution:** Optimized local performance by implementing custom SQLite indexing and integrating `ripgrep` for blazing-fast initial file discovery.
**Challenge:** Preventing "context bloat" in persistent memory.
**Solution:** Built a semantic deduplication layer that calculates cosine similarity on ingestion, updating existing nodes rather than creating redundant vectors.
03the harvest
1. **Persistence:** Achieved 100% cross-session recall of architectural decisions, eliminating the "stateless assistant" bottleneck.
2. **Precision:** AST-aware chunking improved the signal-to-noise ratio in retrieved context by ~60% compared to standard fixed-size windowing.
3. **Performance:** Sub-100ms semantic search on codebases with 10k+ files through optimized local vector indexing.
4. **Privacy:** Successfully delivered production-grade AI intelligence with zero external data leakage, fulfilling the highest enterprise security standards.