Agentic memory
The question: How should an AI agent remember what it learned in past conversations, so it does not start from zero every time?
Agents that forget everything between conversations repeat the same investigations. This journey is about giving them a memory that stays useful as it grows.
This page sums up the journey so far. The full write-ups are the entries below. Start with Entry 1: Agent Memory Meets Graph: Introducing MemGraph — Long-Term Memory for AI Cloud Agents (13 min), or read the newest, Entry 2: From a memory graph to a folder of files: why we retired MemGraph (6 min).
Entries (the full posts)
- Agent Memory Meets Graph: Introducing MemGraph — Long-Term Memory for AI Cloud Agents 2026-03-18, start, 13 min, 4.7k tokens
How MemGraph stores what agents learn as a graph of typed memories, curates it after every conversation, and ranks it by what has proven useful. - From a memory graph to a folder of files: why we retired MemGraph 2026-09-25, change of mind, 6 min, 1.6k tokens
Why we replaced MemGraph's typed graph and LLM evolution pipeline with markdown files that the agent reads and edits itself, and what went wrong on the way.
What we think now
Revision 2, 2026-09-25. Older revisions: r1.
- Memory is a folder of markdown files in the agent's sandbox. The agent reads and edits them with its own file tools, in the same turn as the work. (entry 2)
- The agent curates at write time. It reads the index and the leaf, then merges into the leaf. No background LLM rewrites memory. (entry 2)
- One index, MEMORY.md, reaches the prompt. It lists at most 20 tasks with a status. The leaves are read on demand. (entry 2)
- Vector search stays only for small atomic observations. Code enforces only what a file instruction cannot, such as the path guard and the read-time rules of typed subtrees. (entry 2)
Where we changed our minds
- 2026-09-25
Store memories as a graph of typed nodes and edges, curated after each conversation by an LLM evolution pipeline.Store memories as markdown files that the agent in the conversation writes and curates itself, with one index in the prompt. (entry 2)
Open questions
- At what size does a workspace's memory folder need ranked search again?
- How do we catch drift in rules that only the prompt enforces, such as the 20-task index and "edit, do not append"?
- What should happen to a memory that keeps being contradicted?
Follow this journey: RSS. Plain text version: /journeys/agentic-memory.md