Anthropic's Claude Code documentation states it plainly: each session begins with a fresh context window, and CLAUDE.md files and auto memory are the two mechanisms that carry knowledge across sessions. Other assistants work the same way in principle. A decision the assistant made yesterday is gone unless it was written somewhere that gets loaded or searched today.
Keep a decision log the assistant can read
Record the date, the decision, the reason and what was rejected. The rejected options matter: without them, a new session will happily propose the approach you already ruled out. Keep each entry short and link to the longer document rather than pasting it.
In Claude Code, import the log from CLAUDE.md with an @docs/decisions.md line so it loads each session, or reference it in a path-scoped rule. Keep in mind that imports still consume context, and Anthropic suggests keeping CLAUDE.md under about 200 lines; move older decisions to an archive file once they are settled.
docs/decisions.md
## 2026-10-28 Use Postgres row-level security for tenancy
Why: one schema, audited in one place.
Rejected: schema-per-tenant (migration cost).
Source: docs/adr/007-tenancy.mdUse the memory your assistant already has
Tell the assistant explicitly when something is a decision: "Record this in decisions.md" or "remember that we chose X because Y." Then start a new session and ask what was decided, to confirm it was saved where you expect.
| Assistant | Built-in mechanism | What to watch |
|---|---|---|
| Claude Code | CLAUDE.md files plus auto memory in ~/.claude/projects/<project>/memory/ | Only the first 200 lines or 25KB of MEMORY.md load at session start |
| GitHub Copilot | Repository custom instructions plus Copilot Memory (public preview) | Unused memory entries are deleted after 28 days |
| ChatGPT | Memory managed under Settings > Personalization | Memory is personal, not a shared project record |
When decisions must cross tools or teammates
Instruction files live in one repository and are read by the tools that support them. If you also discuss the same project in ChatGPT or Claude Desktop, or several people need the same record, a shared memory store reached over MCP avoids keeping several copies in sync. Options include Basic Memory (Markdown notes), Mem0 OpenMemory, Supermemory and Central Brain.
Central Brain stores decision, fact, question, todo and summary notes plus a project brief as readable local files. The memory_checkpoint tool appends a note, and on connect central_brain_ping returns a RESUME block from saved project memory, so a new session in a supported client starts with the current decisions. Claude Desktop, Claude Code, GPT Desktop, GitHub Copilot in VS Code, Hermes and OpenClaw, connected over standard MCP with a configuration the app generates for each client. If you only use one assistant in one repository, a decision log in the repo is simpler and enough.
Common questions
Why does my assistant forget a decision I told it yesterday?
The decision stayed in the old conversation. Each new session starts with a fresh context window and only loads saved instructions, memory or files it retrieves.
Should decisions go in CLAUDE.md itself?
Keep rules and short current decisions there, and link a separate decision log for history and rationale so the always-loaded file stays short.
Does Central Brain update my decisions.md automatically?
No. A checkpoint writes Central Brain project memory. Keep any separate Markdown log deliberately, and resolve conflicts between them.