Most MCP memory servers store text. Few tag where it came from.

By Matthew Firth, founder of ModelBrain · Sep 14, 2026 · 3 min read

Look across the MCP memory servers people actually run and almost all of them share one property: once something is saved, it is just "a memory." Here is our own proposal: compare memory servers by whether storage records how each memory got there. Of the six we checked against their own READMEs, only Mem0 and ModelBrain tag anything about origin, and neither separates evidence from summary yet.

Six servers, compared

Read from each project's own README and docs as of publish time; check their repos directly before relying on specifics. server-memory · mcp-memory-service · Mem0 · Letta · Zep

The distinction that gets lost

Three things can end up in a memory store, and they carry very different weight: something you told the assistant directly, something the assistant read from a document you gave it, and something the assistant itself concluded or summarized from other memories. Stored as undifferentiated text, all three come back identically confident. A model-generated summary that drifted slightly from the source looks exactly as trustworthy as the sentence you typed yourself, because nothing in the data says otherwise.

Why it matters specifically when a memory is wrong

This stops being an abstract data-modeling concern the moment a recalled memory turns out to be stale or incorrect. Which of the three types it was determines what the right fix even looks like: correcting a stated fact you'd change your mind about is different from re-reading a source document that changed, which is different again from a summary that needs to be regenerated once its inputs move. A memory store that never captured which one it was can't tell you which fix applies; it can only tell you something doesn't match anymore. What a forget call actually has to guarantee covers the failure modes that follow from exactly this gap.

How ModelBrain does it instead

Every memory ModelBrain stores records where it came from, you or a connected document, rather than collapsing everything into one opaque blob. A finer scheme that also separates source evidence, your own assertions and model-generated summaries is designed, not yet built. How ModelBrain tags what it remembers has the actual mechanics.