An open-source persistent memory layer for AI coding agents that stores and retrieves project decisions and context using Postgres and pgvector, aiming to reduce context window size and improve agent consistency.
I’ve been working on an open-source project that acts as a persistent memory layer for AI coding agents. The basic idea: current LLMs are useful, but they forget project decisions, repeat mistakes, drift across sessions, and often require dumping huge amounts of context back into the prompt. This project stores durable memories in a database, embeds them, and retrieves only the relevant ones based on the current prompt/task. The goal is to keep context windows smaller while giving tools like Copilot, Claude, Cursor, local LLMs, etc. access to the same long-term project memory. At its core, it uses Postgres + pgvector to store and search memories. The memory layer tracks things like project decisions, user preferences, prior fixes, architecture notes, corrections, and lessons learned. The LLM doesn’t “own” the memory — the database does. Embeddings are used as semantic pointers so the agent can retrieve only what matters instead of dragging an entire chat history into context. I’ve also been experimenting with evaluator loops around the memory system: deciding whether retrieved context is sufficient, checking whether a response may drift from known project context, and tracing what memories were used or updated. The larger goal is to make coding agents less amnesiac and less dependent on one giant context window. This is still early/prototype-stage, but I’m considering pushing it further as a local-first open-source tool for developers using AI coding assistants. I’d be interested in feedback from people building with AI agents, RAG, MCP, local LLMs, or coding-agent workflows. Specifically: * Does this solve a problem you’ve run into? * Would you use a local persistent memory layer for coding agents? * What integrations would matter most: MCP, VS Code extension, Cursor, Claude Code, CLI, API? * What would make you trust or not trust an AI memory system? The long-term idea is simple: bring your own LLM, but keep your own persistent brain. see link below
agentmemory is an open-source persistent memory layer for AI coding agents (Claude Code, Cursor, Gemini CLI, Codex CLI, etc.) that uses knowledge graphs, confidence scoring, and hybrid search to give agents long-term memory across sessions via MCP, hooks, or REST API. Built on the iii engine, it requires no external databases and exposes 51 MCP tools.
PROJECTMEM is an open-source, local-first memory and judgment layer for AI coding agents that records development events and provides deterministic warnings before repeating failed actions, reducing token waste and improving reproducibility.
The author introduces an open-source MCP server running on Cloudflare Workers that provides persistent, searchable memory for AI clients like Claude, ChatGPT, and Cursor using vector embeddings and duplicate detection.
The author introduces Noosphere, an open-source memory and wiki layer for AI agents and humans, seeking feedback on its structured Postgres-backed memory, Redis-cached recall, and human-editable wiki pages.
A developer discusses limitations in current AI agent memory systems and proposes a new memory layer tool with episode storage and replay debugging, seeking community validation.