LabscoMCP early access ↗
alexandrulecaMCP SERVER

ImprintMCP

Local semantic memory for coding assistants — decisions, patterns and fixes that survive the session.

by alexandruleca ↗Agent MemoryCommunity

LABSCO SUMMARY

The hooks are what keep the memory current.

Manual saving is where memory tools usually die, so this attaches two hooks to Claude Code: a Stop hook that parses each finished exchange for decisions and stores them, and a PreCompact hook that forces a save before the context window is compressed — the moment context is otherwise lost. Everything from embeddings to the vector store runs locally, so the ongoing cost is disk and CPU rather than credits.

What it is

A memory layer that runs entirely on your machine: EmbeddingGemma-300M through ONNX for embeddings, a Qdrant daemon on 127.0.0.1:6333 for vectors, Chonkie for chunking, and SQLite for a temporal fact graph. Your assistant searches memory instead of re-reading files, and the same store is shared across Claude Code, Cursor, Codex CLI, Copilot and Cline. Nothing leaves the box unless you turn on the optional cloud tagger.

What you get

  • search runs semantic search and hydrates the top results by confidence, with filters for project, type, lang, layer, kind and domain
  • store saves a decision, pattern, finding, bug, architecture note or milestone; embedding and tagging run in the background
  • neighbors finds cross-project kin of a memory, and graph_scope navigates the knowledge base by project, topic, source or chunk
  • list_sources, file_summary and file_chunks list indexed files and pull a specific chunk range
  • ingest_url and ingest_content bring in a web page or an inline blob, deduplicated by name
  • kg_query and kg_edit read and change temporal facts, including marking one no longer valid
  • status reports memory count, active facts and a per-project breakdown

Requirements

One install script on Linux, macOS or Windows, then imprint setup <target> wires it into each coding tool you have — imprint setup all does every one that is installed and skips the rest with a warning. Docker images and an imprint update command are provided; updates preserve data/ and the virtualenv. No account and no API key by default. Cloud LLM tagging is opt-in through imprint config set tagger.llm true, with Anthropic, OpenAI, Gemini or a local Ollama or vLLM behind it.

Setup effort

One command — curl -fsSL https://raw.githubusercontent.com/alexandruleca/imprint-memory-layer/main/install.sh | bash

See installation options