v2.1.0-alpha — Architecture Complete

First public release of MemForge

Neuroscience-inspired memory system for AI agents with sleep cycles that actively rewrite and strengthen stored knowledge.

What's built

• Tiered memory (hot → warm → cold) with PostgreSQL, pgvector, and Redis
• Hybrid search — keyword FTS, semantic vectors, reciprocal rank fusion
• Knowledge graph — entities and relationships via recursive CTEs (no Neo4j)
• Sleep cycles — 5-phase background processor: scoring → triage → LLM revision → graph maintenance → reflection
• Memory revision engine — LLM rewrites low-confidence memories (augment, correct, merge, compress)
• Reflection & meta-reflection — LLM synthesizes insights, detects contradictions, extracts procedures
• Outcome feedback — track whether retrieved memories led to good outcomes
• Active recall — proactively surface memories before agent actions
• MCP server — 17 tools for Claude Code, Cursor, and MCP-compatible AI tools
• TypeScript SDK — with ResilientMemForgeClient for graceful degradation
• LLM tool definitions — OpenAI function calling and Anthropic tool_use formats

Status: Alpha

Architecture is complete. Code compiles and passes type-checking. Not yet validated against a live database. We're releasing early because the design is novel and we want feedback.

Get Started

See INTEGRATION.md for how to wire MemForge into your agent (any framework, any language).

See ROADMAP.md for the long-term vision.