MindGardener

Local-first long-term memory for autonomous agents. Built for OpenClaw. Creates wiki knowledge graph from conversations, scores events by surprise, detects c...

MindGardener is a local file-backed memory tool that matches its description, but review these before installing: 1) Confirm you want a CLI that will create and write to workspace/memory/, memory/entities/, graph.jsonl and MEMORY.md (these are the main artifacts). 2) If you don't want external LLM calls, set provider to a local model (Ollama) in garden.yaml — otherwise you'll need to export API keys (GEMINI_API_KEY/OPENAI_API_KEY/ANTHROPIC_API_KEY) yourself. 3) Do not copy the example nightly-cycle.sh verbatim: it references absolute paths (/root/clawd) and an external auto-recall.py; adapt paths and inspect any referenced scripts. 4) If you enable multi-agent sync or integrations that post to channels (#state-tracker, discord, etc.), audit the sync/posting code paths to confirm where data is transmitted and to which endpoints. 5) Because the package writes human-readable markdown files, you can inspect and version-control (git) the memory directory; test in an isolated workspace first if you want to audit behavior before using with real data.

Type: OpenClaw Skill Name: mindgardener Version: 1.1.0 The MindGardener (also referred to as Engram) skill bundle is a sophisticated long-term memory system for AI agents that uses local Markdown files and knowledge graphs. It implements advanced features like 'surprise scoring' (prediction error) to consolidate memories, temporal decay, and identity-level belief tracking. The code is well-structured, modular, and includes an extensive test suite. All high-risk behaviors, such as network calls to LLM providers (Google, OpenAI, Anthropic) and local file manipulation, are strictly aligned with the stated purpose of extracting and retrieving agent memories. No evidence of data exfiltration, backdoors, or malicious prompt injection was found.