Remnic helps agents understand the people they work with: their preferences, projects, constraints, decisions, patterns, and definition of good, so they can help with fewer unnecessary questions.
Technical version
Remnic is an open-source memory and context layer for user-aware agents, focused on scoped memory, provenance, retrieval quality, user correction, and ask-versus-act decisions.
Local-first storage is a trust feature. The category is user-aware agents: systems that need to understand a user over time without turning personalization into uncontrolled surveillance.
Most memory tools store facts. Remnic models working context: what the agent needs to understand about this user to act well right now.
Retrieved memory should carry enough metadata for the agent and user to understand why it appeared and whether it belongs in the current context.
Personalization without boundaries becomes surveillance. Personalization with correction and control becomes agency.
A good agent should spend the user's attention carefully. Remnic can help decide whether an agent should ask, draft, act, refuse, or escalate.
Memory evals
Remnic treats memory quality as something to measure. The eval surface should answer whether memory reduced repeated context, reduced unnecessary clarification, retrieved the right memory, avoided stale memory harm, respected scope, asked when it should have asked, acted when it had enough context, and improved the final output.
Demos to build toward
MCP server, ChatGPT Apps-compatible demo, memory inspection UI, "why did you remember this?" UI, and forget/correct/scope controls.
Buyer memory for brand preferences, size and fit, budget thresholds, purchase constraints, excluded products, gift preferences, shipping urgency, and ask-before-checkout rules.
Repo conventions, preferred architecture patterns, test expectations, release process, past bugs, review preferences, and ask-before-changing-public-API rules.