Enterprise AI Memory

Air-Gapped Agent
Memory

Strict 420MB offline sidecar. Drops noise, extracts facts. Zero API calls, no context rot.

HIPAA SOC2 Air-Gapped On-Premise
Request Architecture Review
mnemostroma SYSTEM ARCHITECTURE · INFRASTRUCTURE TOPOLOGY REV 2026.04 · STRICT CONDUCTOR PROXY TCP / HTTP Intercept Dynamic Context Injection stdio · Daemon INJECT <memorycontext> L1 · I/O BOUNDARY CLIENT (IDE/CLI) User Interface Standard I/O AI AGENT LLM Engine FORK STREAM L2 · ASYNC SIDECAR PIPELINE OBSERVER (NON-BLOCKING CAPTURE) Total: ~20ms latency Regex Pass 0.1ms · PII strip HybridNER 10ms · Entity Extr. e5-small 15ms · Vector Embed TinyBERT Ranker 6ms · Core logic L4 · DATA ROUTING & DISPATCH "THALAMUS" ORCHESTRATOR (PERSISTENCE HUB) Queue Batching Parsed Kernels L3 · MEMORY CORE (HETEROGENEOUS) RAM INDEX In-Memory Vector Engine fp16·512d Relevance Fn R·T·I Score Retrieval Latency ~20ms p95 Eviction Policy LRU + Decay Session Bound 200-500 cap SQLITE WAL Disk / IO Core Ledger Fact Vault Anchor Constraints Decision Trees Write Mode Async Append Storage Engine WAL2 + fsync Hydration Cold Boot Hot Vectors Fsync Facts L5 · DECAY ENGINE (BACKGROUND CRON) DREAMER BACKGROUND WORKER Trigger: Idle 5m Dissolution Pipeline: FULL TEXT → GIST → SKELETON EMBEDDING Compression Ratio: ~85% reduction. Evicts low LRU-score records from RAM Index. Scan low-score L6 · MCP API SURFACE MCP EXPOSED TOOLS (READ-ONLY) Transport: stdio / JSON-RPC ctx_semantic() ctx_anchors() ctx_bridge() ctx_search() Agent sends query args Return Vectors Return Facts Data Payload (JSON to Agent) MAX FOOTPRINT 600MB RAM INFRA DEPENDENCY ZERO (Local SQLite & ONNX) RUNTIME ENGINE ONNX Runtime INT8 / CPU

Vector DB Trap vs Active Forgetting

Logs pile up → drift + OOM. Mnemostroma: Background process extracts facts to SQLite, noise decays.

Problem

Context Rot

Every session adds noise. Vectors drift. Costs explode.

Solution

Active Dissolution

Signals decay biologically. Core facts persist. Pure efficiency.

System Mechanics

Three layers. One sidecar. Zero agent changes.

1

Passive Observer Sidecar

Zero agent changes. 80% token savings via selective extraction.

2

Biological Decay

Old noise fades. Semantic core persists forever.

3

Pure Context Mode

Agent autonomy. XML facts only, no instructions.

System Constraints

420MB
Baseline
20ms
p95 Retrieval
3x
ONNX INT8
SQLite
WAL

Architectural Review

30min technical call. Quantify your token waste and constraints fit.

Schedule Review