Home › How memory works
What Kagura remembers
Kagura's memory isn't a chat log, and it isn't a pile of document chunks. It stores your work as structured memories — decisions, deadlines, places, numbers, know-how — that stay organized, stay current, and come back exactly when you need them.
Memory has more than one shape
Ask the questions you'd ask a teammate with a great memory. Each one is something Kagura is built to remember.
“Why did we decide it this way?”
Decisions & reasoning — The conclusion is saved together with the why — the options weighed, the trade-offs, who signed off. The next session starts from the decision instead of re-litigating it.
“What's due next week?”
Dates & deadlines — A memory can carry a date or a time window. Ask what's coming up, and plans, deadlines, and scheduled re-checks surface right on time.
“What did we do at this site last time?”
Places — A memory can carry a location. On site or on the move, pull up what happened nearby — past visits, incidents, decisions made on the spot.
“How much has this changed since last month?”
Numbers over time — Costs, weights, scores — any number can be recorded as a series, so you can ask for the trend, the average, or just the latest value.
“How did we get around that bug?”
Know-how & workarounds — Troubles and their fixes become searchable experience. The same trap never costs your team twice.
Memory that grows
Saved memories don't just sit in a list. Related entries link into a web of context, duplicates merge, and stale or contradicting notes get cleaned up over time — automatically on Pro with Sleep Maintenance. The more you use it, the more accurate your knowledge base becomes.
Short-term memory, long-term memory
Some memories are working notes; others are meant to last. Both are recallable the moment they're saved. On Pro, Sleep Maintenance periodically reviews your short-term memories, promotes the ones worth keeping to long-term memory, and archives the ones that have served their purpose. Archiving isn't deletion — archived memories can be restored for a period of time.
Facts change — history stays
The facts of your work change: the deploy target moves, an owner rotates, a policy gets revised. In Kagura, a new memory can be recorded as replacing an older one. Everyday recall returns only the current fact, while the old memory stays on as history — so you can still trace what used to be true, and why it changed.
When two memories disagree and it isn't yet clear which is right, Kagura marks them as a contradiction and shows both; neither side is silently hidden. Kagura may also suggest a replacement, but accepting or dismissing it is always your call.
Memory that's always on — and memory you recall
Most memories come back when you search for them. But some things should never need searching: goals, standing rules, guardrails. Pin those, and every session starts with them already loaded — while everything else stays quietly out of the way until you ask.
Pinned memories are treated as long-term right away and are exempt from automatic cleanup. However much merging and archiving goes on, a pinned rule is never removed automatically. Memories that carry a date are protected the same way.
Remembered safely
Your memories are private by default; you decide what becomes shared with your team. Memories that arrive through connectors like Slack are summarized with personal data redacted before anything is stored, and they carry their origin with them — so your AI can always tell first-hand knowledge from external chatter.
Under the hood
The memory engine is open source. Architecture, the hybrid search design, the MCP tool reference, and self-hosting docs all live on GitHub.
Technical docs on GitHub →Try it yourself
About ten minutes takes you from first sign-in to your first remembered decision.