Memory Now Remembers How Things Changed — A Version History For What Alfrada Knows About You

July 5, 2026

Until now, when a fact about you changed, the old one was quietly overwritten and the reasoning behind it was gone. Memory is now a version-controlled timeline: every change Alfrada makes to what it knows about you is a commit with a readable diff, old facts are superseded rather than deleted (kept and marked "true until…"), and people, projects, and places get their own lane so you can watch a relationship or a plan evolve over time. Alfrada can also re-mine your entire chat history to rebuild that timeline as if it had been tracking you all along.

What you can do

• Read your memory like a git log — open Digital Twin → Memories and every change appears as a dated commit: a one-line summary, who made it (you, the agent, a migration, or a re-mine), and a colour-coded tally — + created, ~ superseded, − invalidated, ± edited. Expand any commit to see the exact before/after lines.
• See facts change instead of vanish — when something updates ("moved from Barclays to Wise"), the old fact is superseded: kept in history, struck through, and stamped with the date it stopped being true, while the new fact takes over. Nothing is silently lost.
• Trace one fact's whole life — click any fact to open its Fact history drawer: the full version chain from first mention to current truth, each version's dates, category, entity tags, and origin, with edit and delete inline on the live version.
• Browse by person, project, or place — switch to the Entities view to get a lane per tagged entity (person / org / project / place). Pick "John" or "Alfrada" and see every fact about them ordered by when it was true, active facts in green and retired ones struck through.
• Rebuild your timeline from your real history — a background re-mine walks your archived conversations in chronological order and re-extracts memories anchored at each conversation's own date, so a fact from March is dated March, not today. It reconstructs real episodes, supersede chains, and first-seen dates. A banner shows live progress (12/40 conversations); it's resumable and picks up where it stopped if interrupted.
• Add a memory by hand, with time — the Add memory form takes the fact, its type (life story / rule / world info), an optional topic chip, and an optional "when did this happen?" date so manual entries land on the timeline in the right place.
• Get time-aware recall in chat — retrieval now pulls facts whose event date sits near today's window, not just when they were written, so date-anchored context ("this happened last July") surfaces at the right moment.

Where this shows up

• You told Alfrada in March you bank with Barclays, then switched to Wise in June. You used to end up with a contradiction or a lost fact. Now the Barclays fact reads "true until June", the Wise fact is current, and the switch is one clean step on your timeline.
• You want to know why Alfrada assumed something about you. Instead of guessing, you open the fact, see its full version history — what it replaced, when, and whether you or the agent wrote it — and correct it in place.
• You've had hundreds of conversations before this landed. Rather than starting the timeline from today, Alfrada re-mines your whole history in the background and hands you a memory graph that already knows when each thing happened.

Try it

• "What do you remember about my move from Barclays to Wise, and when did each change?"
• "Show me everything you know about the Alfrada project, in order."
• "That fact about my role is out of date — here's the correction."
• "Rebuild my memory timeline from my past conversations."

Heads up

• Automatic upgrade, nothing lost. Existing memories are imported as dated "genesis" commits, so your history starts populated. The pipeline moves to v2 (bitemporal entries, an append-only event log, episodes, and entity tags) on its own — no action needed.
• Superseded ≠ deleted. Retired facts stay in history so the timeline stays honest; explicit delete still removes a fact when you actually want it gone.
• Re-mine runs quietly. It starts in the background after service boot, paces itself between conversations so it never starves live chat, and is billing-exempt — the historical re-extraction does not spend your credits. Normal in-conversation memory extraction is billed as before.
• Entity tagging is best-effort. Facts are tagged with named entities by a light extra pass; if it can't tag something, the memory still saves — tags are enrichment, never a gate.