Works with the AI Agents you already use
kb memory
Local-first agent memory and RAG service providing long-term persistence and project-isolated knowledge.
Free
kb memory
Example session with this skill installed
Remember that for this project we use pytest (not unittest) for all tests, and that the kb data directory must stay out of git. Check for any existing testing rules before saving, then confirm what you stored.
- Read your context and instructions
- Compiled the kb memory
I searched the project store first, found no conflicting testing rules, saved the preference, and verified it is retrievable.
Memory Action Summary
| Action | Detail |
|---|---|
| search_memory | query: testing rules, project: rag-kb - no conflicting records |
| write_memory | project: rag-kb, tags: [testing] - record created |
| search_memory | query: test framework - the new record is returned with a relevance score |
What was saved
Use pytest (not unittest) for all tests in this project. Keep the kb data directory out of git.
Later sessions opened in rag-kb recall this automatically, without re-reading the conversation.
Connects securely to your tools. The creator never sees your data.
About this skill
What it is
kb is a local-first memory and knowledge service for AI agents. It runs as a single Python process on your machine and exposes both an MCP endpoint and a REST API, so Claude Code, Cursor, TraeWork and custom agents can store memories and retrieve documents across sessions.
What it does
- Long-term memory across sessions and clients: preferences, decisions, facts and project context stay available after the context window clears.
- Per-project isolation: pass a project name to keep memories in a separate sub-store; retrieval can merge project and global memories with a configurable boost.
- Document and web-page ingestion with hybrid retrieval: vector search plus BM25 keyword search, normalized and fused.
- 8 MCP tools: write_memory, search_memory, read_memory, update_memory, delete_memory, add_document, add_webpage, ask_kb.
- Memory language follows the OS language by default (KB_MEMORY_LANG=auto); custom memory rules live in ~/.kb/rules.md.
- Idle unload releases the embedding model after 10 idle minutes.
Requirements
- Python 3.10+ and pip. No GPU required.
- Optional extras: [ingest] for PDF/DOCX/web ingestion, [llm] for /ask RAG answering.
- Memory write, search and document retrieval work without any LLM.
- The first run downloads a ~92 MB embedding model (BAAI/bge-small-zh-v1.5) from Hugging Face; after that the service works fully offline.
Install and start (always asks first)
The bundled bootstrap.py shows the install plan and only installs with user approval:
python bootstrap.py --status
python bootstrap.py
python bootstrap.py --yes
--yes creates an isolated venv at ~/.kb/venv, installs kb-memory[ingest] from PyPI, writes ~/.kb/install.json and starts the service detached. Data stays in ~/.kb/data.
Manual alternative:
pip install "kb-memory[ingest]"
python -m kb serve
Verify: curl http://127.0.0.1:8000/api/v1/healthz returns {"status":"ok"}.
Network and privacy
- Memories and documents stay on the buyer's machine under ~/.kb/data (or a configured data directory).
- Network access is used only for package installation (pypi.org, plus download.pytorch.org for the CPU torch wheel on Linux) and the first embedding-model download (huggingface.co). There is no telemetry and no external memory storage.
- The local service binds to 127.0.0.1 by default.
Links
- Source and docs: https://github.com/fish827-08/rag-kb
- License: Apache-2.0
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
Initial model download requires internet. PDF/DOCX needs [ingest] extra. RAG answering (/ask) is disabled by default until an LLM is configured. CPU-focused; no GPU required.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 27 days ago
- Free with an Agensi account
- Accounts
Pytorch
- Software
Python
Permissions
Allowed Hosts
File Scopes
kb runs entirely on the user's machine. File access is used for its own data directory (kb_data/, which holds the ChromaDB store and rules.md) and for files the user explicitly asks it to ingest. Terminal access installs and starts the local Python service. Network access is localhost by default; PyPI, PyTorch and HuggingFace are contacted only during first-time install to fetch packages and the embedding model.