Part of #204 (Open Loop). Rom: systematically use special files in the user repo as a persistence layer + AI memory (DECISIONS.md, CODE-OVERVIEW.md, KNOWLEDGE-BASE.md, BRAINSTORMING.md, ...).
We already write DECISIONS.md (the decisions ledger). This extends the idea to the read side: the agent reads these files at the start of a run for context, and is told to keep them current as it works.
First slice (framework, deterministic on our side):
- loadRepoMemory(cwd) reads the canonical memory files present in the workspace.
- runFramework frames them into the system prompt: current contents (so the agent starts with that context = memory) + an instruction to maintain them (create if missing = persistence). The agent has file access, so it does the writing.
- CLI reads repo memory from the workspace and passes it in (same pattern as signals + the-framework.yml). Empty repo just gets the maintain instruction.
Follow-ups (separate): a summarize/update step we run on the ledger, and any structured schema for the files.
Part of #204 (Open Loop). Rom: systematically use special files in the user repo as a persistence layer + AI memory (DECISIONS.md, CODE-OVERVIEW.md, KNOWLEDGE-BASE.md, BRAINSTORMING.md, ...).
We already write DECISIONS.md (the decisions ledger). This extends the idea to the read side: the agent reads these files at the start of a run for context, and is told to keep them current as it works.
First slice (framework, deterministic on our side):
Follow-ups (separate): a summarize/update step we run on the ledger, and any structured schema for the files.