Why Your AI Coding Agents Suffer from Amnesia (And How I Fixed It with SQLite + ONNX)

MShesat — Thu, 29 Oct 2026 16:50:39 +0000

If you use Claude Desktop, Cursor, or OpenCode, you know the daily struggle: every new session is a blank slate. Standard RAG solutions treat memory as flat vector chunks, which makes them fundamentally time-blind.

To solve this, I built Synapse 4D, an open-source local-first cognitive memory engine that acts as a permanent hippocampus for AI agents.

Core Architecture

D1 Focus: Working memory buffer with strict token management.
D2 Distiller: Extracts atomic facts and resolves contradictions.
D3 Chronos: Bi-temporal SQLite database (valid_from, valid_to) with Hebbian recency decay.
D4 Mesh: Reciprocal graph ensuring zero orphan stubs.

🧠 Permanent AI Memory: Integrates natively with Claude Desktop & Cursor via MCP.
⏳ Bi-temporal Tracking: Resolves contradictory facts without erasing historical context.
🔒 100% Local & Private: Pure SQLite + ONNX embeddings. Zero cloud APIs, zero telemetry, zero PyTorch bloat.

Check the open-source repository on GitHub:
👉 https://github.com/MShesat/SYNAPSE-4D

For those interested in a visual 3D neural graph HUD and a Windows installer, I also packaged a Pro Edition:
👉 https://synapse4d.lemonsqueezy.com/buy

DEV Community: MShesat

Why Your AI Coding Agents Suffer from Amnesia (And How I Fixed It with SQLite + ONNX)

Core Architecture