"kgai vs mempalace: verbatim transcripts or a decision log"
mempalace and kgai agree on a surprising amount. Local-first, MIT, built because AI coding sessions forget. The disagreement is the unit of memory. mempalace keeps what was said. kgai keeps what was decided. We build kgai, so read this knowing that.
What mempalace is good at. It stores your conversation history as verbatim text and retrieves it with semantic search. Wings, rooms and drawers give people, projects and topics real structure, and retrieval is scoped instead of run against one flat corpus. Hooks capture Claude Code, Codex and Cursor transcripts before context compression eats them. It publishes strong recall numbers on LongMemEval with zero API calls at retrieval time. If the thing you need back is the actual exchange, in the actual words, this is a tool that takes that seriously.
Where kgai differs, structurally.
- One log, everything else is a projection. mempalace maintains a vector store and a temporal knowledge graph side by side. An open mempalace issue (#224) describes the failure mode when the two diverge, facts invalidated in the graph still surfacing through vector search, and calls it a memory correctness risk. kgai stores one append-only log. The graph is rebuilt from it, so there's no second layer to fall out of step.
- Supersession is explicit. A new decision supersedes an old one through a link with a reason. Recall returns only what's in force, and the full chain, dead ends included, is one command away. That model is the subject of what decision memory is.
- No daemon. Writes go through the CLI into per-writer shards. Nothing holds a lock, nothing runs in the background.
- Team sync goes through an S3 bucket you own, once you turn it on, and since no two writers share a shard there's nothing that can conflict textually. mempalace's README doesn't address team sharing.
On speed, one number we'll stand behind. A decision lookup answers in about 100 ms at 1M decisions across 30 writers' shards. Recall and free-text search are slower than that.
Choose mempalace if you want verbatim recall of what was said, across every agent you use, with semantic search over it. Choose kgai if what your team keeps losing is why the code ended up this way. kgai deliberately doesn't keep conversations and doesn't personalize per user. If you're comparing local-first options more broadly, kgai vs engram is the nearer fight.
Installing takes two commands and leaves nothing behind but plain files in the repo. Run it beside whatever you use today, they don't compete for the same memory.
Try it on your repo. Two commands, MIT, nothing leaves your machine until you configure a bucket.
See how kgai compares to other tools on the comparison page, or browse the other notes.