Cognikernel

AI coding assistants forget everything between sessions. CogniKernel addresses this by attaching to Claude Code and Codex through their hook surfaces, pulling out the decisions, constraints, and abandoned approaches worth preserving, and consolidating them into an event-sourced store. On the next session, that store injects a compact context block — the agent arrives knowing what was already decided. The store is keyed on the project path, so the same memory file travels between tools working on the same project without any manual export or sync.

The core differentiator is what CogniKernel is not. It does not wrap a vector database, does not call an embeddings API, and does not require a running server. The vendor describes it as an event-sourced log of typed memos — structured records with a defined schema, not free-form text blobs. That structure is what makes the log auditable: you can read, inspect, and version-control exactly what the agent is being told about your project’s history.

This fits well on solo developer setups and small teams that want zero-cost, fully local memory without standing up infrastructure. Self-hosting is the only deployment mode, which is either a feature or a constraint depending on your organization’s appetite for managing local tooling. The gap shows when a project needs memory that goes beyond typed decisions — semantic search over past discussions, retrieval ranked by relevance, or any form of reasoning about what the codebase means rather than what was decided about it. For those needs, a vector store or a retrieval-augmented generation layer is still required alongside or instead of this tool.

The project is Apache-2.0 licensed and implemented in Python, with the package structure indicating a src layout under the memlora module. It includes a test suite and contribution guidelines, suggesting it is structured for external contribution, though community adoption at time of writing is early-stage.