Read this before you rely on it.

Preview: not independently security reviewed.

  • The API and on-disk format may change before v1.0.
  • Tamper-evident, not tamper-proof. Anyone who can write the files can rebuild the hash chain. Keep the head fingerprint that Verify() returns somewhere else and compare it later.
  • Isolation stops mistakes and prompt injection, not the host program. Whoever holds the store or the folder can read every agent’s files.
  • Over MCP the model is not trusted, but it is powerful. It declares the sources and legs itself, and can recall or erase anything in its agent. A model that can run shell commands can bypass the tools and read every agent’s files.
  • Legs are declared by the caller. If a memory was really learned from a source but no leg says so, recalling that source won’t remove it.
  • Only Erase removes text from disk. Untrusted hides memories until you run Compact(). Backups and copies keep their data.
  • The embedding server sees what is remembered and searched. A local one keeps it on your machine; a hosted one keeps whatever its provider keeps, and an Erase cannot reach that copy.
  • Search is exact, and the command reloads it every time. Every question is compared with every vector: no missed results, and each searchable memory takes about 0.8 KB of RAM (768 values at one byte each). But the command and MCP server read the whole index again on every search: 9 ms at 1,000 memories, 0.74 s at 100,000, 7.5 s at 1M (numbers).
  • Forget finds memories, not sources. Sources are not in the search index; recall a source by its number.
  • Damage near the end of an active log is cut off and reported in Report.DroppedBytes, so check it after opening.

The full list is in the README and the threat model.

Use something else when

  • You need SQL-style queries, filters or keyword search. CheeseDB finds by id or by meaning.
  • Startup time matters with very long histories: opening an agent replays its log, about 0.7 s at 1M memories. Snapshots are planned.
  • An agent’s memory won’t fit in RAM: about 205 bytes per memory, plus about 0.8 KB per searchable memory.
  • You need production guarantees today.

Measured numbers, including where CheeseDB loses to SQLite, are on the numbers page.