Membox: memory that survives your agents
A memory box for AI agents — the missing piece between stateless chat sessions and a real working memory: store, recall and reuse what your agents learn across runs.
Agents have a goldfish problem: every session starts from zero. Context windows fill up, sessions end, and the hard-won knowledge — which config worked, which endpoint moved, why that one test is flaky — evaporates.
Membox is my take on the missing middle layer: a small, deliberate memory box that agents can write to and read from across runs.
The shape of it
- Boxes, not dumps. Memories are short, labeled entries — a name, the fact, and why it matters — not an endless transcript to re-summarize.
- Recall beats storage. The point isn't remembering everything; it's surfacing the right entry at the right moment, so an agent starts a task already knowing the last task's lessons.
- Boring tech. The box stays queryable and diffable — memory you can audit beats memory you can't.
Where it fits with this portal
The same idea runs through Z-Assist's self-learning: staff-approved answers re-enter the knowledge base, so the chat gets smarter instead of just older. Membox generalizes that instinct — treat agent knowledge as a first-class artifact, version it, and let every new run inherit it.
(Building in public: this one lives outside the portal — notes here will grow as it matures.)