oleksiijko/pmb

Local-first persistent memory for AI coding agents (Claude Code, Cursor, Codex) over MCP. Decisions, lessons and facts live in one SQLite file on your disk. Offline, multilingual.

What it solves

AI coding agents often forget critical project decisions, lessons, and constraints between sessions, forcing users to re-explain the same context repeatedly. PMB provides a local-first, durable memory system that allows agents to recall past decisions, project structures, and personal facts without relying on on-cloud storage or API keys.

How it works

PMB operates as an MCP (Model Context Protocol) server that connects to agents like Claude Code, Cursor, and VS Code. It uses a hybrid retrieval system combining BM25 (lexical), dense vectors (semantic), and an entity graph to fetch relevant context.

Key technical components include:

  • Storage: SQLite serves as the source of truth for durable events, while LanceDB handles rebuildable vector indexes.
  • Read Path: A prepare(message) call retrieves project context, lessons, and recent activity in 4-16ms, injecting it into the agent's prompt before the model processes the request.
  • Write Path: Memory is captured asynchronously (under 1ms) via agent tools or "ambient memory" hooks that journal the agent's work (edits, tests, commits) even if the agent forgets to explicitly record a fact.
  • Multilingual Support: Uses a multilingual embedder covering 50+ languages, allowing cross-lingual retrieval (e.g., a Russian query finding an English fact).

Who it’s for

Developers using MCP-aware AI coding agents (such as Claude Code, Cursor, Codex, Windsurf, or Zed) who want their agents to maintain long-term project memory across sessions and model upgrades while keeping their data entirely local.

Highlights

  • Local-first: No cloud, no API keys, and no telemetry; all data is stored in a local SQLite workspace.
  • Protocol-level Hooks: Injects memory automatically before the model thinks, removing the need for the agent to manually call a memory tool.
  • Ambient Memory: Automatically journals agent activity (like successful fixes or deploys) to ensure real work is captured.
  • Visual Dashboard: A local web UI providing a live entity graph (Map), a journal (Timeline), and tools to prune "dead" lessons.
  • High Performance: Extremely fast recall (35ms warm) and near-instant writes.
  • Data Portability: Supports snapshots and exports to Markdown/JSON.

Related

  • Project
    basicmachines-co/basic-memory

    A persistent, two-way memory system for AI assistants that uses shared Markdown files and the Model Context Protocol (MCP) to create a long-term knowledge graph.

  • Project
    24kchengYe/MemoMind

    MemoMind is a locally hosted, GPU‑accelerated memory system for AI coding agents. It stores extracted facts from chats, documents, and daily life events in PostgreSQL + pgvector, builds a knowledge graph, and offers fast 4‑way hybrid retrieval (semantic, BM25, graph, temporal). The agent can retain new information, recall relevant memories, and reflect across the whole store. All data stays on the user’s machine, works with any OpenAI‑compatible LLM, and includes a web dashboard for browsing and exporting memories.

  • Project
    justrach/codedb

    A high-performance context engine built in Zig that provides AI coding agents with a structural map of a codebase via MCP to improve navigation and retrieval.

  • Project
    akitaonrails/ai-memory

    ai-memory is a long-term memory system for AI coding agents, storing memory as git-backed Markdown, working across agents, projects, machines, and teams, with zero-LLM default and optional LLM features.

  • Dispatch
    MCP-Memory: Fast Agent Memory with OKF and SQLite FTS5

    MCP-Memory is an OKF‑compliant server that gives AI agents persistent, searchable long‑term memory backed by SQLite FTS5, enabling sub‑20 ms lookups and human‑readable markdown storage.