GitHub Repository

Neuroscience-inspired memory for AI agents — sleep cycles that rewrite and strengthen knowledge over time. TypeScript, PostgreSQL, MCP.

14 starsTypeScript

Self-improving agent memory system, 92% R 5 LongMemEval, PostgreSQL

Name: Self-improving agent memory system, 92% R 5 LongMemEval, PostgreSQL
Availability: InStock
Author: artificium

by artificium·Apr 9, 2026·4 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainNiche Gem

Sleep cycles that actively rewrite memories beat passive vector stores like LangChain memory.

Strengths

•Single PostgreSQL handles vectors, graph, and storage—no Pinecone or Neo4j dependency.
•Memory revision engine measures its own quality via contradiction rates and stability metrics.
•LongMemEval benchmarks (92% R@5) with 8 security audits shows production thinking.

Weaknesses

•Agent memory is a crowded category with LangChain, LlamaIndex, and dozens of vector store solutions.
•Sleep cycle concept is interesting but unclear if it outperforms simpler consolidation strategies.

Post Description

MemForge is an experiment in a single database (PostgreSQL) with local embeddings. The goal is to enable long-term, persistent memory independent from the model or agent framework. It began as an attempt to maintain context over longer periods without the token cost spiraling out of control and the lost context after consolidation.

The architecture is based on a rough simulation of human memory and uses sleep cycle (agent inactivity) where an LLM reorganizes and stores memory in a multi-tier database with a number of sorting, ordering, and prioritization mechanisms. I also attempted to build this with security built in from the beginning.

Since I first published this, it has evolved quite a bit based on some feedback and integrating some ideas from many of the other fantastic memory management systems that show up daily. Hopefully my contribution to the community can help move things forward and bring some value to people as a full solution, or just to carry some ideas forward in other projects.

Looking for feedback on how the sleep cycle works, the bet on a PostgreSQL only solution - particularly from anyone willing to run this under real workloads. Full transparency -- it's Alpha now, but has undergone multiple rounds of stress testing, hardening, and benchmarking.