Long-term memory for your agents
What was decided, tried and learned, kept in Git between sessions and readable by every agent you run.
Who it’s for: Developers running custom agents, and teams using several assistants on the same work.
Agents forget between sessions
Each new session starts without what the last one learned. Teams end up repeating context in prompts, or building memory into each agent separately.
Memory as files in a repository
Write what your agents learn to Markdown and commit it, by hand or with your own tooling. Notedigit serves that repository read-only, so every agent can search and read it at the start of a task.
Custom agents connect with a token
Assistants such as Claude and ChatGPT sign in over OAuth. Custom agents can connect with a token, and every call they make lands in the audit log with the tool used and the note read.
How to structure the repository
- Keep an index note listing what each memory file covers. An agent can read it first and go straight to the right file.
- Record failures as well as decisions: what was tried, what happened, and why it was dropped. That is the part that gets repeated most often.
- Write memory in plain, dated sentences. The same files stay readable to the people reviewing what the agents know.
Setting it up
-
Step 1: Create a memory repository and commit your agents’ output, or your own notes, to it.
-
Step 2: Connect it in Notedigit. Give custom agents a token and people their own accounts.
-
Step 3: Tell each agent to search memory at the start of a task. Every read shows up in the audit log.