Memory - Open Source, agent-maintained wiki for coding agents
Hey IH, I’m Michele, one of the creators of Memory:
https://github.com/aictx/memory
Memory is an MIT-licensed open-source project for giving coding agents repo-specific context before they start work.
The problem we’re trying to solve: coding agents usually rediscover the repository every time. They can inspect code, but they often miss the durable project knowledge: product intent, architecture decisions, setup/debugging workflows, repo conventions, known gotchas, abandoned approaches, constraints, and open questions.
Memory is not chat history. The goal is to preserve project knowledge that future agents should not have to rediscover.
It stores context locally under .aictx/ as structured objects with provenance and relations. Core memory commands do not require embeddings or a model API, so the stored context can be inspected, diffed, reviewed, and versioned with the repo.
There is also a local viewer, demo via:
https://demo.aictx.dev/
The viewer lets humans browse the same memory objects, source records, relations, provenance, schema, and graph context that agents load. One of our design goals is that agent-maintained memory should be reviewable by default, not a hidden pile of notes.
A few use cases we’re building toward:
1. Keeping AGENTS.md, CLAUDE.md, Cursor rules, and similar files small by treating them as operating manuals, while durable project knowledge lives in local memory.
2. Loading task-shaped context before coding, debugging, reviewing, architecture work, or onboarding, so agents start from relevant decisions, workflows, gotchas, and constraints.
3. Using the local viewer as a project-state dashboard for agent-built projects: what the product is, what stack was chosen, what decisions are source-backed, what questions are still open, and what the next agent can safely assume.
The project is still early. The core workflow works, but many hard parts are unresolved: ranking, recall, save timing, avoiding memory noise, schema design, repo-size defaults, evals on real projects, and integrations with agent tools.
Docs:
Project site:
We’d especially like feedback from people using coding agents on larger or long-lived repos:
What project context is actually worth preserving? What should never be saved? And how would you want agent-maintained memory to stay reviewable?
