wtf is engram-lite?

engram-lite is a local memory system for AI agents, meaning a tool that lets AI assistants remember facts from earlier conversations and keep that knowledge after the program restarts. Instead of giving every AI agent the same pile of memories for the same question, engram-lite serves each agent only the memories that fit who it is, based on a profile you register for it that describes its persona, its subject area, and the kind of task it handles. The README explains that feeding an agent one shared pile of memory actually caused it to make things up more often than giving it no memory at all, while serving memory this targeted way roughly doubled how often agents succeeded at their task and cut down false statements to about a third of what they were before. The whole thing runs on one machine with no cloud service, no account, and no external server. All memories are stored in a single SQLite database file, plus a local embedding model used to search that memory by meaning rather than just exact keywords. It cleans itself up over time by skipping unimportant input like plain commands, removing duplicate facts, letting newer facts replace older outdated ones, and expiring facts that go stale. Every decision the system makes about keeping, dropping, or merging a memory is logged with the reason behind it, so a user can ask why something was or was not remembered and get a real answer. Installation is done with a normal pip install, and there is a small Python library for reading and writing memories directly in code. The project also ships ready made integrations for three specific tools: a Hermes AI agent memory provider, a Claude Code plugin that hooks into prompts automatically with the user's consent, and an OpenClaw plugin, all of which can share one memory store while still keeping each agent's results scoped to what it is allowed to see. The README also describes a benchmark result on a long conversation memory test called LoCoMo, where the project scores well without making any calls to a large language model to build its memory, keeping that part of the process free to run. The project is written in Python.