Graphs, LLM wikis, gBrain, RAG — that is the data layer.
Memory is a different thing, and an agent needs both kept apart.
Jaegyu Lee · Q00
Ouroboros maintainer
The data layer stores what is true. The memory layer decides what to do with it,
shaped to the input in front of the agent right now.
More data in context is not more competence.
An agent drowning in your mailbox is still a new hire — it reads everything and concludes nothing.
The problem is not access to data. It is the absence of a layer that turns data into a decision.
To do good work an agent needs to read the situation and commit to a conclusion —
matched to the input, not to a fixed script or a pile of retrieved chunks.
Good for guardrails. But a fixed path can’t bend to a situation it didn’t anticipate.
Brings the right chunks. Still hands you raw material, not a decision.
Reads this input against past judgment and decides what to do next, here, now.
The frontier labs have already moved past retrieval.
The open question is no longer “what do we store” but “how does an agent remember.”
Building memory systems shaped to pull harder on FOMO — engagement as the design target.
Saving why the data was made, to make recall efficient — a vision for agentic memory beyond the developer’s lens.
In most stacks the data layer and the agent layer are not cleanly separated yet.
An enterprise agent must know the caller’s authority before it answers.
A line employee can ask the CEO agent a confidential question — the response and the file search must be scoped to their permission.
Permission is not a wrapper around the agent. It is a gate inside every retrieval and every answer.
CRDT solves humans editing the same doc at the same time.
Agents don’t need that. They need to produce clean markdown and HTML, with permission management on top. That’s the whole spec.
The hippocampus runs a memory advisory; long-term holds a theta profile.
A time axis tells the agent how it grew — injected as first-class memory so it knows the shortcut.
The time axis is the point: the agent carries who it became, not just what it once saw.
Over roughly a day of context data, the agent reached the same outcome with 1.74 fewer tool calls on average — because the shortcut was already in memory.
Fewer calls is not just speed. It is the agent knowing where to look before it looks.
The data layer stays one shared, queryable source of truth.
The memory layer is personal, situational, and decides — that is the line between a new hire and a senior.
Don’t bolt memory onto your data. Build it as its own layer.