It's been 13 years since my last post here, but I figured this might be of some interest. I know there's a LOT of AI slop around, I hope you'll not consider this that.
Background: I left Augment, due to their pricing rug pull, and then missed their excellent context engine. I wanted to replicate it, but instead ended up creating something, i think, a touch more useful.
This is a 'holographic' memory system for AI (agents / all-types / code / openclaw etc). It is a crude representation of the human mind. It references (among others) https://github.com/WujiangXu/A-mem-sysm, vector symbolic architectures & the Zettelkasten method etc
There are 2 parts:
Fold: "holographic" storage (based on some interesting research from the 80s), semantic search & auto-indexing engine. Stores and indexes everything (code, docs, notes). Works with git or local fs. Indexer + some pretty interesting vector/embedding/llm goodness.
Engram: an active memory agent that writes experience and concepts as it goes, dropping them into Fold to be related to other memories.
• 5-tier memory model inspired by human cognition: Core -> Conscious -> Subconscious -> Short-term -> Fold (infinite semantic search)
• Two modes: Personal (emotional tagging, soul/identity) and Work (shared project memory, hit counts only)
• Hit tracking promotes frequently-referenced memories up tiers, FIFO evicts stale ones
• "Sleep" cycle for consolidation (like human sleep; promotes, demotes, merges etc)
• Works with Claude Code, OpenClaw, or any agent framework with sub-agent support
• File-based (markdown), no database lock-in (rebuilds from fs)
• MIT licensed, fully local
• Separation of concerns: Fold = raw knowledge, Engram = lived experience
• Agent-readable installer (point your agent at AGENTS.md (http://agents.md/), it self-installs)
My issue is that i'm working on VERY large code bases. Fold is very good at breaking down the initial 'recover context' stage into much shorter times. Instead of spending the first 10 minutes of every session re-explaining my project, the agent already knows what we were working on, why we made certain decisions, and what we plan to do next. It picks up where we left off.
We've made it open-source obviously, are are thinking we 'might' spend a little time to make a paid hosted version, not entirely sure. Any thoughts are welcome ofc.
Hope you all are having a lovely day or night.
0 comments
—
Show HN: Not another sloppy memory tool · HackerTrans
Background: I left Augment, due to their pricing rug pull, and then missed their excellent context engine. I wanted to replicate it, but instead ended up creating something, i think, a touch more useful.
This is a 'holographic' memory system for AI (agents / all-types / code / openclaw etc). It is a crude representation of the human mind. It references (among others) https://github.com/WujiangXu/A-mem-sysm, vector symbolic architectures & the Zettelkasten method etc
There are 2 parts:
Fold: "holographic" storage (based on some interesting research from the 80s), semantic search & auto-indexing engine. Stores and indexes everything (code, docs, notes). Works with git or local fs. Indexer + some pretty interesting vector/embedding/llm goodness.
Engram: an active memory agent that writes experience and concepts as it goes, dropping them into Fold to be related to other memories.
• 5-tier memory model inspired by human cognition: Core -> Conscious -> Subconscious -> Short-term -> Fold (infinite semantic search) • Two modes: Personal (emotional tagging, soul/identity) and Work (shared project memory, hit counts only) • Hit tracking promotes frequently-referenced memories up tiers, FIFO evicts stale ones • "Sleep" cycle for consolidation (like human sleep; promotes, demotes, merges etc) • Works with Claude Code, OpenClaw, or any agent framework with sub-agent support
• File-based (markdown), no database lock-in (rebuilds from fs) • MIT licensed, fully local • Separation of concerns: Fold = raw knowledge, Engram = lived experience • Agent-readable installer (point your agent at AGENTS.md (http://agents.md/), it self-installs)
My issue is that i'm working on VERY large code bases. Fold is very good at breaking down the initial 'recover context' stage into much shorter times. Instead of spending the first 10 minutes of every session re-explaining my project, the agent already knows what we were working on, why we made certain decisions, and what we plan to do next. It picks up where we left off.
We've made it open-source obviously, are are thinking we 'might' spend a little time to make a paid hosted version, not entirely sure. Any thoughts are welcome ofc.
Hope you all are having a lovely day or night.