# Why I Built Zekra: Shared AI Agent Memory, on ToGO

> I run agents in Claude Code, Codex and Cursor, and each session started from zero. This is why I built Zekra, and why on ToGO.

*2026-10-19T12:49:04Z*

# Why I Built Zekra: Shared AI Agent Memory, on ToGO

I run a lot of AI agents across Claude Code, Codex and Cursor. Every new session started from zero, and I kept re-explaining the same decisions. That is the whole reason Zekra exists.

![Why I built Zekra](/uploads/sites/3b93f650-28ec-4093-99aa-e7b16c2a2e2f/media/why-i-built-zekra-og.png)

## What I wanted

One memory that every agent could reach, without custom glue for each tool. MCP was the obvious door: if an agent speaks MCP, it can use the memory.

I also did not want a pile of notes. Every write in Zekra becomes an add, an update, an invalidate or a no-op, so a changed decision replaces the old one instead of sitting next to it. Recall mixes vectors, BM25 and a knowledge graph, and it works in Arabic and English.

## How it started

The repository history shows the order of work. The first commits are from 16 July 2026: the source-of-truth artifacts, then a ToGO app scaffold, then a split into a thin app and a separate `brain` plugin that holds the memory logic. Later commits add the TEI embedding plugin, hybrid BM25 recall, a Redis cache and single-binary serving. The project began under the working name cabrain; the legacy `CABRAIN_*` environment variables are still honoured.

As of this writing the repo has 350 commits on `main`, counted with `git rev-list --count origin/main`.

## Why on ToGO

ToGO already had the database, cache, queue and storage pieces as plugins. That let me put the memory logic in its own plugin, with its own module and tests, and keep the app around it small. The full details, with code and counts, are in the [case study](https://to-go.dev/en/blog/zekra-togo-case-study), including what was awkward.

## Why open source

Zekra is open source under AGPL-3.0, so you can self-host it. There is also a hosted version at app.zekra.dev. I use it every day, and my own assistant runs on it.

## What I would like to hear

What would you want your agent to remember first, and what should it never keep? I am asking for feedback, not votes.

## FAQ

**Why did you build Zekra?** Because every agent session started from zero and I was tired of repeating decisions.

**Why build it on ToGO?** The infrastructure plugins already existed, so the memory logic could be its own plugin.

**Is it open source?** Yes, under AGPL-3.0-only.

## Links

- Product article: [Self-hosted AI agent memory with Zekra](https://zekra.dev/en/blog/self-hosted-ai-agent-memory-server)
- Case study: [Zekra on ToGO](https://to-go.dev/en/blog/zekra-togo-case-study)
- GitHub: [fadymondy/zekra](https://github.com/fadymondy/zekra)
- Product Hunt: [Zekra](https://www.producthunt.com/products/zekra)
- Site: [zekra.dev](https://zekra.dev)

---

Source: https://fadymondy.com/en/blog/why-i-built-zekra
