Skip to main content
A memory layer is defined as much by what it refuses to keep as by what it stores. This page is the honest version of both.

​
Keep vs. never keep

Worth keeping

  • Decisions — what was chosen, and why
  • Repo rules — conventions the work must hold to
  • Plans — the shape of work in flight
  • Durable facts about the project or the people on it

Never stored

  • Secrets — keys, tokens, credentials
  • Personal data — redacted before write
  • Transcripts, tool calls, file contents
  • Anything a newer decision has retired
The test for whether something belongs: would the next agent be worse off without it? A decision passes. A log line of a file being read does not.

​
Redaction fails closed

Secrets and personal data are stripped before anything is persisted — not on the way out. If content cannot be made safe, the write is rejected rather than stored in a degraded form.
This is a backstop, not a licence. Redaction is pattern-based, and no pattern set catches everything. Do not deliberately route credentials or customer PII through memory tools on the assumption that they will be caught. Prompt your agents not to send them in the first place.

​
Superseding

When a newer memory contradicts an older one, the older one is superseded and stops surfacing in recall. The chain is retained — the history of what was believed when is not deleted — but only current truth comes back from a query. This is what keeps recall sharp over months. Without it, a store accumulates contradictions and the agent picks between them arbitrarily. To retract something explicitly rather than waiting for it to be contradicted, use forget with the memory id from a prior recall.

​
Deduplication

Near-identical writes are collapsed rather than stored twice. Agents are repetitive by nature — several of them working the same project will write the same conclusion in slightly different words, and without dedup the store fills with paraphrases of a single fact.
Dedup is similarity-based, so it is a judgement call rather than an exact match. Two genuinely distinct facts phrased almost identically can be collapsed. If you need both, make the wording meaningfully different or scope them to different spaces.

​
Where it physically lives

Memory lives on Klio’s infrastructure, encrypted in transit and at rest at the infrastructure level, isolated per org. Secrets and PII are redacted before storage.The encryption keys are ours, and Cloud writes are not hash-chained. If your threat model requires keys you hold, self-host.
The user-held key and the hash chain are properties of the self-hosted engine specifically. Anywhere you see those claims, they refer to self-hosting — Klio Cloud makes the weaker guarantees described above.