Launched this week
Memorist
An auditable, local-first memory engine for Open WebUI
7 followers
An auditable, local-first memory engine for Open WebUI
7 followers
Long-running LLM work loses continuity: decisions, corrections, and constraints evaporate between sessions. Most fixes summarize old chats into a hidden prompt — invisible, unauditable, easy to poison. Memorist takes a different position: raw messages are evidence, not memory. Every message is preserved unchanged in a local ledger. Deterministic sentence segmentation produces units with exact character offsets, so every memory claim traces back to the exact words that produced it.
Persistent memory feels like foundational infrastructure for AI. Curious—what design tradeoff ended up being much harder than you expected while building Memorist?
@aryan787544
Hello. First of all, I would like to thank you for your attention to this product.
It was deciding what the system is allowed to believe about the user.
Look, at first glance, such a project might seem like something that stores a series of important facts and recalls them as needed.
But the main point is that in the design and engineering of the pipelines, we had to consider when and based on what criteria, the memory engine should decide what to remember in a typical chat, with what degree of importance and with what relevance; even more complicated, when and how to decide to forget something or close a file in its mind? And this guideline gave shape and direction to many designs in the memory engine. Finally, the pipeline became:
conversation → signals → candidates → validation/gating → memory authority → retrieval
A memory candidate is no longer a fact, but rather an object that must be determined and acted upon by the processing nodes:
provenance (where did this come from?)
confidence
creation context
user control state
lifecycle
possible contradiction
ability to be revoked
@cna_farokhi Thanks for the thoughtful explanation.
I'd rather not take up more space in the Product Hunt thread.
If you're open to continuing the conversation, what's the best email to reach you?
@aryan787544 Always happy to talk about these things, you can contact me at: cna.farokhi [at] gmail [dot] com
@cna_farokhi Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.
Coming soon: Local Claude Code integration plugin.