Memory That Knows It Might Be Wrong

A memory layer for AI agents that tags every fact with source, confidence and freshness — and says so when it isn't sure.

Day 0

Teach it something

Each fact is tagged with who it's about (scope), what it's about (key), and how it was learned (source) — that tag is what lets the engine reason about confidence later.

Quick facts to try:

Ask it something

Retrieval always shows its confidence — and flags itself when the answer might be wrong.

Ask a question to see what the memory layer surfaces.

Memory inspector

"What do you remember about me?" — a user-facing view of everything currently active, with a working forget button per fact.

Architecture snapshot: write path → retrieval path → forgetting path

Write path

new fact → slot lookup by (scope, key) → contradiction? → supersede old + store new

Retrieval path

query → match key/scope → compute confidence from source + decay → flag "might be wrong" if warranted → answer + reasons

Forgetting path

tick sweep (stale) · new write (contradicted) · explicit request (revoked) → status changed, audit logged, excluded from retrieval

Activity & audit log

Every write, contradiction, decay-forget, revocation and reconfirmation — timestamped by day. This is the exportable compliance record.