To onboard agentic colleagues into existing teams we need them to continually improve themselves, by learning from the real world, reflecting on those memories and adapting its behaviour according to insights from reflecting.
Self-improvement requires
- agent being involved in actual collaboration and/or agent has access to human knowledge bases to generate raw material
- its own persistent memory
- a mechanism to reason and reflect on memories regularly and distill knowledge or surface insights
Approaches
There are several approaches to this.
- Karpathy’s AI Librarian has been the most influential. It uses a knowledge-base-in-context approach where the system maintains a working memory of relevant information and retrieves it as needed. Karpathy later expanded on the idea with LLM Wiki. A very early version can be seen in The “append-and-review note”.
- Marciuo Puga’s Cog, a prototypical “plain-text cognitive architecture for Claude Code”. (I also like his blogging (Think in Public) while developing it in a separate blog only for Cog.
- GitAgent approaches this from the repo-side: defining agents in Git repos, including memory and reflection.
- Brenno Ferrari’s obsidian-mind, the same basic idea but simpler and clearly scoped to an Obsidian-based agent.
- Claude Code has a built-in memory management (“Auto Memory”) and the 2026 Claude Code source leak (news post, analysis), confirmed the idea: 3-layer design and Claude’s internal “autoDream”: background memory rewriting.
- Piskala et al.: From Everything-is-a-File to Files-Are-All-You-Need: How Unix Philosophy Informs the Design of Agentic AI Systems explains how the principles of simplicity and composability, enshrined in the UNIX philosophy give us established tools and processes to interact with text.
- classic architecture with LLM, RAG and MCP