MemoryGate_Standalone

MemoryGate v1.0.0 - First Production Release
Forgetting That Remembers: Intelligent Memory Retention for AI Systems
This release introduces MemoryGate's forgetting mechanism - a production-ready system for managing AI memory through retention scoring, tier-based archival, and audit trails.
What's New
Retention Scoring System

Memories gain score when accessed (+0.4 per retrieval)
Score decays over time (-0.02 every 15 minutes)
Access patterns naturally surface important content

Tier-Based Archival

HOT tier: Actively used memories (score ≥ -1.0)
COLD tier: Archived summaries (score < -1.0)
Automatic migration based on access patterns

Soft Deletion with Audit Trails

Tombstone records preserve deletion metadata
Configurable purge thresholds (default: score < -2.0)
Full causality tracking for compliance

Technical Highlights

Migration 0002_cold_storage: Adds retention columns to all memory tables
Backward compatible: Existing data migrates seamlessly with default scores
Configurable parameters: Adjust bump/decay rates via environment variables
Production tested: Zero errors across 15 MCP tool integrations

Deployment
Production instance running at memorygate.fly.dev with PostgreSQL + pgvector backend. Schema revision 0002_cold_storage applied successfully.
Breaking Changes
None. Fully backward compatible with existing clients.

Retention Parameters:

SCORE_BUMP_ALPHA=0.4 (access boost)
SCORE_DECAY_BETA=0.02 (time decay)
SUMMARY_TRIGGER_SCORE=-1.0 (archive threshold)
PURGE_TRIGGER_SCORE=-2.0 (purge threshold)
RETENTION_TICK_SECONDS=900 (15 min cycles)