I started building CogniKernel to solve this problem.
Instead of relying on markdown memory files, CogniKernel observes the coding session itself. It captures decisions, constraints, architectural decisions, project specific context etc info as they naturally emerge.
The goal is simple: The agent should remember what it learned yesterday without me telling it what to remember.
Some things it stores automatically:
- Architectural decisions and design rationale - Project conventions and coding patterns - Important implementation details - User preferences and recurring constraints - Relationships between files, modules, and components
The memory is local-first and designed to work alongside existing coding agents rather than replace them.
A few design goals I had:
- Zero manual memory curation - Local storage instead of cloud dependency - Fast retrieval with relevant context instead of dumping everything - Agent-agnostic architecture so it can work beyond a single coding assistant - Memory that improves over time instead of growing into an unsearchable log
Was it impactful??? I build 4 full fledged projects using cognikernel while switching the auto-memory off from claude code and same project same prompt, same sessions using the auto memory
- 30% Less Tokens used - 4x less read tools used
And a bonus!!! If you use claude code and codex both you can just start a session in alternate tool and your memory will carry forward there as well!!!
This was a hobby project and if you use it and it helped your work, please let me know any feedback!