Persistent project memory
PMB records decisions, lessons, facts, files, and entities into a single local memory store so the agent can reuse them later.
PMB
PMB is a local-first memory layer for AI coding agents, storing decisions, lessons, and facts in a SQLite file on your disk. Works with Claude Code, Cursor, Codex, and Zed.
Local memory for AI coding agents
PMB is a local-first memory layer for AI coding agents. It gives tools such as Claude Code, Cursor, Codex, and Zed a shared place to store decisions, lessons, facts, and project context in one SQLite file on your disk.
The product is designed to inject relevant memory before an agent answers and to write new information after the turn, so context persists across sessions without API keys, cloud services, or telemetry. The source also describes a dashboard for exploring the memory graph and timeline, plus support for running optional summarization and extraction locally.
Core capabilities
Pre-answer recall and post-action journaling
Hooks inject relevant memory before the model answers, and PMB journals the agent’s work afterward, so recall does not depend on the agent remembering to ask for it.
Hybrid retrieval
The source says recall combines BM25, dense vectors, an entity graph, and optional reranking, then fuses the results with reciprocal-rank fusion.
Non-blocking writes
Writes are asynchronous and the tool returns immediately; the write path is described as sub-millisecond while indexing continues in the background.
Dashboard for inspection and exploration
PMB surfaces a live map and timeline for exploring memory by entity, kind, cluster, and event order.
MCP-based integration
The product integrates over MCP using stdio, with the server running as a child process of the agent and no network port or token required.
Practical workflows
Project memory across sessions
Use PMB to keep project conventions, decisions, and bug history available across sessions so a fresh chat does not start from zero.
Cross-agent continuity
Use one shared workspace memory when switching between Claude Code, Cursor, Codex, or Zed so context follows the project instead of the editor.
Workflow with memory hygiene
Let PMB capture lessons and rules during day-to-day coding, then revisit the dashboard to see which memories are being followed and which ones should be pruned.
Exploring stored context
Use the local dashboard’s map and timeline to inspect what was recorded, how items connect, and what happened most recently in a project’s history.
Offline local-model workflows
Run PMB with local models for summarization or graph extraction when you want the memory pipeline to stay offline end to end.
Pros and Cons
Pros
- Keeps memory local in a single SQLite file on disk.
- Works offline and does not require API keys on the read path.
- Supports multiple coding agents from the same workspace memory.
- Uses hybrid retrieval rather than a single lookup method.
- Provides a dashboard with map and timeline views for inspecting stored context.
- Can run optional model-based passes on local models such as Ollama.
Cons
- Pricing details are not available from the provided pages.
- The documentation captured here is partial, so some setup and integration details are still unclear.
FAQ
Does my code or data leave my machine?
PMB runs as a local SQLite-backed memory store on your machine. The source says there are no network calls on the read path, no account, and no telemetry, and that it still works offline.
How is PMB different from a vector database?
The source describes PMB as using hybrid recall: BM25, dense vectors, an entity graph, and optional reranking, fused with reciprocal-rank fusion. It is presented as a memory layer rather than a plain vector database.
Which coding agents does PMB work with?
PMB supports Claude Code, Cursor, Codex, and Zed according to the homepage. It connects over MCP/stdio and can be wired into an agent from the terminal.
Can PMB run fully offline?
Yes. The homepage says you can bring your own model or run the optional summarize and graph-extract passes on a local model such as Ollama.
What is PMB best suited for?
The source says PMB is useful for project memory, decisions, lessons, facts, and keeping agent context across sessions and tools. It is aimed at coding workflows where the same project context should persist.
Quick Facts
- Category
- Developer Tool
- Primary use
- Persistent memory for AI coding agents
- Platform
- Local-first; macOS, Linux, and Windows are mentioned in the quickstart
- Storage
- SQLite file on disk, with LanceDB next to it for vectors
- Integrations
- Claude Code, Cursor, Codex, and Zed
- Source domain
- pmbai.dev
PMB Alternatives
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
Ghost
Ghost is a terminal-based AI assistant for chatting, code generation, and CLI tasks. Includes free models, supports Linux, macOS, Windows, and is open source.
Goldfish
Goldfish is an AI memory app for macOS that helps you reply, write, summarize, and continue with context already on your Mac. It stores memory locally and writes in your tone.
Lasso
Lasso is an ecommerce product data platform for enriching catalog records, processing supplier files, generating product content, and tracking competitors. It includes a web app, REST API, SDK, and MCP server.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
Struere
Struere turns spreadsheet data into structured operational software with dashboards, alerts, and automations for teams replacing manual spreadsheet workflows.