Some time ago I wrote here about the missing memory: the way today's AI assistants remember like a logbook. Everything is stored, everything weighs the same, and what gets retrieved is governed by recency or similarity. A coffee break sits next to a life-changing decision, without ranking.
EWMC — Emotionally-Weighted Memory Consolidation — is my attempt to take that seriously. The idea is simple, and borrowed from how we remember ourselves: we remember what we feel and what turns out to have consequences, not everything we process. Each memory gets a weight that changes over time. What returns and proves meaningful persists; trivia fades; and a nightly consolidation pass cleans and reorganizes the store, much as a night's sleep sorts the impressions of a day.
I am now releasing it as open source, in two parts.
The first repo, ewmc, is the memory store itself: a small, local reference implementation. Everything runs on your own machine — one SQLite file and a local Ollama model for embeddings. No cloud, no account, no keys. Your memories never leave the computer.
The second, local-ai-memory-os, is a pattern on top: how several assistants — say a chat client and a coding agent on another machine — can share one durable memory through a git-distributed log, where anything that becomes core memory always passes human review first.
I want to be honest about what this is and is not. It is not new fundamental research. The ingredients are familiar — consolidation as metaphor, frequency and consequence weighting, forgetting curves and repetition, vector memory. The contribution is the combination, and a couple of opinionated choices: negative memories fade faster (to avoid rumination loops), and nothing is promoted to core memory automatically — that decision is yours.
The code
- ewmc — the memory store
- local-ai-memory-os — orchestration across several AIs
- The companion text: EWMC white paper
It is free to use, change and build on. I would be glad to hear from you if you try it.