i killed my baby and it was the best decision i ever made
only a few thousand of you saw my CAM post, the 10,000 line semantic memory interface with embeddings and knowledge graphs and claude hooks.
i found after about a week of using it:
- it worked
- slow </3
what actually happened
spent some time building this elaborate memory infrastructure. vector db. sqlite. semantic search. auto-ingestion pipelines. relationship graphs. the whole nine yards.
it worked! claude remembered stuff! problem solved right?
except...
every session took 4+ seconds just to boot. try running 6 ghostty sessions with a pretty big chunk of filled context windows. i was basically watching claude fibbergitting (aka eating my ram up)
the thing i built to make claude more "performant" was making claude slower.
so i thought:
"am i engineering around claude or working with it?"
refactoring:
threw it all out. started over.
new stack:
- two bash scripts
- global/project CLAUDE.md files
- claude code hooks
- thats it
session starts → context loads from markdown
session ends → state saves to markdown
no api calls. no database. no dependencies.
1,500 lines total.
insight:
agents dont need elaborate memory infrastructure.
they need a persistent layer thats:
- simple enough to trust
- light enough to ignore
- powerful enough to persist
turns out CLAUDE.md files + bash scripts + hooks can do everything the 10k line monster did. just... cleaner. faster. & more maintainable.
the philosophy shift
i stopped trying to build around claude's limitations and started building with claude's strengths.
the original system was me trying to be clever and attempt novelty (thx adhd)
"claude has no memory? ill build a whole ass database!"
the new system is me being smart.
"claude can read markdown and bash is fast as hell. lets just use that."
less infrastructure = less bottlenecks = more windows = more velocity
unsevered memory
thats what im calling it.
same problem solution. 93% less code. 10x faster. actually maintainable.
sometimes the move is subtracting not adding.
sometimes your 10,000 line "solution" was just you over-engineering because you could.
---
tl;dr - rewrote my entire claude memory system. went from 10k lines with databases to 1.5k lines with markdown files. boots instantly now. runs 6 windows without lag. learned that simple beats clever every single time.
link to original severance thread: https://www.reddit.com/r/ClaudeAI/comments/1phtii5/unsevering_claude_to_my_codebase_achieving/ if you wanna see how we got here
link to git repo: https://github.com/blas0/UnseveredMemory
only a few thousand of you saw my CAM post, the 10,000 line semantic memory interface with embeddings and knowledge graphs and claude hooks.
i found after about a week of using it: - it worked - slow </3
what actually happened
spent some time building this elaborate memory infrastructure. vector db. sqlite. semantic search. auto-ingestion pipelines. relationship graphs. the whole nine yards.
it worked! claude remembered stuff! problem solved right?
except...
every session took 4+ seconds just to boot. try running 6 ghostty sessions with a pretty big chunk of filled context windows. i was basically watching claude fibbergitting (aka eating my ram up)
the thing i built to make claude more "performant" was making claude slower.
so i thought:
"am i engineering around claude or working with it?"
refactoring:
threw it all out. started over.
new stack:
- two bash scripts
- global/project CLAUDE.md files
- claude code hooks
- thats it
session starts → context loads from markdown
session ends → state saves to markdown
no api calls. no database. no dependencies.
1,500 lines total.
insight:
agents dont need elaborate memory infrastructure.
they need a persistent layer thats:
- simple enough to trust
- light enough to ignore
- powerful enough to persist
turns out CLAUDE.md files + bash scripts + hooks can do everything the 10k line monster did. just... cleaner. faster. & more maintainable.
the philosophy shift
i stopped trying to build around claude's limitations and started building with claude's strengths.
the original system was me trying to be clever and attempt novelty (thx adhd)
"claude has no memory? ill build a whole ass database!"
the new system is me being smart.
"claude can read markdown and bash is fast as hell. lets just use that."
less infrastructure = less bottlenecks = more windows = more velocity
unsevered memory
thats what im calling it.
same problem solution. 93% less code. 10x faster. actually maintainable.
sometimes the move is subtracting not adding.
sometimes your 10,000 line "solution" was just you over-engineering because you could.
---
tl;dr - rewrote my entire claude memory system. went from 10k lines with databases to 1.5k lines with markdown files. boots instantly now. runs 6 windows without lag. learned that simple beats clever every single time.
link to original severance thread: https://www.reddit.com/r/ClaudeAI/comments/1phtii5/unsevering_claude_to_my_codebase_achieving/ if you wanna see how we got here link to git repo: https://github.com/blas0/UnseveredMemory