Stop Re-explaining Your Code to Brain-dead AI Agents: PMB Brings Local Memory via MCP

Ever feel like using Cursor or Claude Code is like talking to a brilliant intern with extreme ADHD? Every single session, you have to write a mini-novel explaining why your auth layer looks like a sketchy DIY project. If you are tired of this daily "AI amnesia," you aren't alone.

A new open-source tool called PMB (Persistent Memory Bridge) just launched on Product Hunt, scoring over 170 upvotes by promising to give your AI agents a persistent, local-first memory.

The Lowdown on PMB: What’s the Hype?

Created by dev @oleksiijko, PMB is a developer tool that bridges the memory gap for Claude Code, Cursor, Zed, and other MCP-compatible coding assistants.

Here is how this pragmatic tool works under the hood:

• Local-first SQLite storage: It stores all your design decisions, project goals, lessons, and code-entity graphs in a single SQLite workspace directly on your machine. No sending sensitive proprietary code to a third-party VPS cloud.
• Zero-token read path: It does not call an external LLM on the read path, keeping it fast, lightweight, and completely offline-first.
• Dynamic selective retrieval: Instead of dumping an entire bloated .cursorrules file into the prompt, PMB uses a hybrid retrieval engine (BM25 + vectors + entity graph) to inject only the relevant snippets of context needed for the active task.
• Local Dashboard: It includes an interactive dashboard so you can inspect, edit, or purge memories that are no longer accurate.

The Dev Community Speaks: Savior or Just Another Prompt Wrapper?

The Product Hunt launch sparked some highly technical and realistic debates among the community.

The "Shut Up and Take My Money" Crowd: Devs working under strict corporate security guidelines are loving the local-first architecture. It offers the cognitive boost of persistent memory without triggering compliance alarms or requiring cloud APIs.

The "What About Context Drift?" Skeptics: One user asked a killer question: "What happens when architectural decisions change? If week 2's logic contradicts month 6's reality, does the AI start hallucinating outdated patterns?"

The creator jumped in to explain how PMB handles drift:

1. Forgetting-curve decay: Older context naturally loses relevance weighting over time.
2. Corrections override: When you correct your agent, that feedback is saved as a high-priority lesson that immediately outranks older contradicting notes.
3. Keyed facts: Key-value attributes apply a "latest-wins" rule, archiving older data automatically.

The Context Window Worriers: Another concern was that front-loading memories might eat up the LLM's precious context window. The creator reassured them that PMB uses selective top-k retrieval, adding only a few hundred tokens on-demand. In practice, this actually saves more prompt runway than it consumes compared to dumping massive static text files.

The Multi-Client Freelancer Concern: Freelancers working across multiple client repos worried about memory leakage. The creator clarified that isolation is baked into the architecture: each project runs in its own isolated SQLite workspace under ~/.pmb/workspaces/<id>/, making cross-client leakage mathematically impossible.

The C4F Takeaway: Will It Save Your Sanity?

Let’s be real: AI coding assistants are incredible, but their short-term memory is frustratingly limited. PMB is a clever, pragmatic approach that leverages the power of MCP (Model Context Protocol) to put the developer back in control of the context layer.

However, a word of advice from a seasoned dev: no database is completely self-cleaning. If you don't occasionally audit your agent's memory dashboard, you'll eventually find it trying to patch modern bugs with deprecated legacy patterns. Keep your SQLite database tidy, purge the outdated decisions, and enjoy a much smoother AI-assisted coding experience.

You can take it for a spin right now with: pip install pmb-ai && pmb setup

Source: Product Hunt