ContextPool
Persistent memory for AI coding agents
293 followers
Persistent memory for AI coding agents
293 followers
Every AI coding session starts from scratch. You re-debug the same bugs, re-explain decisions you already made. Your agent forgets everything. ContextPool gives your agent persistent memory. It scans your past Cursor and Claude Code sessions, extracts engineering insights (bugs, fixes, design decisions, gotchas), and loads relevant context via MCP at session start. No prompting needed. Works with Claude code, Cursor, Windsurf, and Kiro. Free and open source - team sync available for $7.99/mo.
Built something similar for a different layer persistent memory across business workflows, not just coding sessions. The "docs graveyard" concern from the comments is real. What helped us was making memory write-on-use, not write-on-save. If an agent references a piece of context during a task, that context gets reinforced. If nothing ever pulls it, it decays. Curious how you handle relevance scoring when the pool grows past a few thousand entries.
Vibe-coder here. I maintain a claude.md file and update it manually at the end of every session. It's manual, but it works. For a solo builder (no team) what does ContextPool give me that a well-maintained claude.md doesn't?
probably the extraction part — you document what you think matters,
but i'd guess it catches stuff you'd never think to write down (that
random 2am gotcha that took 3 hours). @majidyusufi curious — what
happens with low-signal sessions, does it still store something or is
there a quality threshold?
ContextPool
@majidyusufi @webappski That 2am gotcha is exactly the sweet spot, the stuff that never makes it into docs because by the time you fix it, you just want to sleep.
On low-signal sessions: there is a quality threshold. Not every session produces a summary worth storing. If a session is just routine edits with no meaningful decisions, patterns, or debugging, it doesn't add noise to the pool.
@majidyusufi @webappski thanks, that makes sense I don't always know what's worth documenting until it breaks again.
runprompt
What I've used so far that works very well for me is the compound part of Compound Engineering. The problem I see to CE is that it's per repo, ContextPool looks amazing since all my repos can share these eng learnings!
Great work!
And what if I had multiple projects in Claude Code? How do you handle that?
UXPin Merge
Interesting concept, but “exhaustive scanning” sounds expensive at scale. Curious how it performs with large document sets in real production use.