AI agents forget everything when the conversation ends. Meridian is the structured notes store our agents read and write so context survives the session, the machine, and the model. It is the memory the rest of the suite is built on.
An AI assistant is brilliant within a conversation and amnesiac between them. Close the window and the context is gone: what was decided, what is in flight, what was tried and rejected. The next session starts cold, and you pay for it by re-briefing.
Meridian fixes that with a deliberately simple idea: a structured store of notes, organised by project and document type, that any agent can read at the start of a session and write back to at the end. State lives outside the conversation, so it outlives it.
Because the store is shared, an answer one agent reaches is available to the next. It is the connective tissue under Poppy, Ganglia, and the rest of what we build.
At the start of a session, an agent lists and reads the notes for the project it is working on — the current state, recent decisions, open work — so it begins in context instead of cold.
It does the task with that context in hand, and can search the wider store whenever it needs a specific fact, runbook, or prior decision.
At the end it records what changed — new state, decisions made, work now in flight — as typed notes under the right project, so the next session inherits it.
The things worth keeping for the long term are promoted out of session traces into durable memory, so a good answer is captured once and reused, not re-derived.
Persistent memory comes in three complementary pieces.
The project-notes store. Typed documents — status, handover, runbook, decisions, knowledge — scoped by project, read and written through a simple command-line interface.
Thought capture. The place a passing idea or insight is recorded, so it is not lost between the moment it occurs and the moment it is useful.
The trace record. A full log of what an agent did in a session, from which the passages worth keeping can be promoted into long-term memory.
Meridian is rarely used directly — it shows up as the reason the other tools feel like they remember you.
The assistant reads and writes its working state here, so it never needs re-briefing. See Poppy →
The thinking system stores each session as a trace and promotes the keepers into long-term memory. See Ganglia →
Any agent we run starts by catching up on Meridian and ends by writing back what changed.
Honest about what this is: Meridian is infrastructure we built for ourselves and run every day. It is not a product for sale. We are showing it because persistent, well-structured memory is what turns a clever chatbot into an agent you can actually rely on — and getting that right is central to what our practice does.