series · agent memory
Agent Memory From Zero
The 51-minute video, written down: six parts, each paired with its own chapter of the video. What we call "memory" is the agent reading the right file — this series builds that sentence from zero.
憶1. What "memory" is — the token problem, RAG and the knowledge graphPart 1/6Memory lives in your folder, not in the agent’s head: remembering is the right file being read. Two paths, three pieces, the RAG detour and what the graph really means.自2. Claude Code's own memory — auto memory and the files behind itPart 2/6Even with nothing installed, Claude Code remembers things. What does it write, where, by which rule — and what gets deleted after 30 days?図3. The LLM Wiki — built on camera with INGEST / QUERY / LINTPart 3/6Vault, index and double brackets: a persistent knowledge graph the agent keeps writing — from zero, with measurements.鏡4. Do you actually need Obsidian? What it does and cannot doPart 4/6The answer is a deliberate no: Obsidian does not build the system, it makes it visible. Graph, backlinks, a pane of glass — and four things it cannot do.帳5. Is NotebookLM enough? Wiring it into Claude CodePart 5/6Why bother when a ready-made tool exists? Where NotebookLM shines, where the wiki does — and the MCP bridge that joins them.使6. Hermes Agent — the agent that manages its own memoryPart 6/6Nous Research’s agent updates its own memory the moment a session ends. A bounded, compressed memory — combined with a growing wiki.