Long story short: all the memory providers kinda suck for different reasons — except one.
Cloud ones suck because they're cloud: vendor lock-in and data retention just isn't for me. Hindsight is technically the best, but too heavy to run — too many API calls, costly even on cheap models, too many bugs. OpenViking is a pain to set up, I dropped halfway. Holographic — liked the speed, quality wasn't there. Hancho — good at profiling, but same problem as Hindsight.
Then I found Mnemosyne. Easiest to set up, lightweight, fully local, best balance of quality and speed. It's not built in by default… but it should be.
Hancho for me — only because it's the one I've actually used. I really don't get why people find these a pain to set up? You tell your agent to set it up and go grab a sandwich. Hardest part is deciding what goes on the sandwich 🥪
Is it really that good? Short answer: yes. Hindsight is SQLite + sqlite-vec for vectors + FTS5 for text. Fully local, zero external deps, MIT license. Genuinely impressed once it's dialled in.
Already running Hindsight in an agent mesh. Thorny road to adoption — finicky — but we nailed it down. Had Opus eval Hindsight vs Mnemosyne vs Signet AI vs agentmemory. Sharing in case it helps anyone.
Tried Hindsight and Mnemosyne. Hindsight has good integration but too many calls, and the rerank pins my VPS at 100% CPU. Mnemosyne does a great job. Ended up writing a native plugin for Graphiti too — very fast so far.
Mnemosyne is great — built a dashboard for it. Another good one is YantrikDB. Gonna check out Signet since people keep mentioning it.
Tried Holographic (a disaster) and now Hindsight, which I find finicky — it silently fails and memories stop processing for days. Once it grows it gets messy. How's Mnemosyne in that regard?
Mnemo is great. My agent set it up by itself.