If a fresh session can't see it, it isn't project context yet
A long agent session can hold the plan, the constraints, and the why. That is conversation context. Project context is what a new session, a new tool, or another engineer can still read.
A long session with a coding agent can leave you with a shared picture: what was decided, what was ruled out, and which constraints still hold. Open a new session, switch tools, or hand the work to a teammate, and that picture is often gone. What remained was the context of one conversation. It was not yet the project's context.
The test is a fresh session
Google's Elephant-Goldfish Model treats this as a design problem, not a prompt-writing problem. In the open-source EGM plugin, the Elephant holds your project context and drafts a specification. Goldfish reviewers examine the draft and its sources without the earlier conversation, so they can catch unstated assumptions.
You do not need to adopt the plugin for the test to be useful. If a session that never had the original conversation cannot reconstruct the current decisions, constraints, and why, you do not yet have project context. You have a chat log.
Context that lives on one machine is not shared
Claude Code's official documentation is direct about where auto memory lives. Claude can write notes about ongoing work, deadlines, and decisions it cannot derive from the code or from git history. Those files sit under ~/.claude/projects/ in the home directory. They are not shared across machines or cloud environments. What you put in the repository can travel with the project. Notes that stay in the home directory cannot.
That gap showed up in practice. In a public feature request, Haustorium12 described building an external memory system after repeated context compactions. Local files stayed machine-specific. A separate vault was synced between a home PC and a work PC, and the setup warned that each machine's local memory was invisible to the other.
Commenting on the same issue, yurukusa wrote that they had arrived at a similar pattern on their own. One lesson they flagged: they used to copy project structure into memory files, and those copies drifted within days. If the code is the source of truth, they argued, do not copy it into memory. They preferred a short briefing — what is done, what is next, what is blocked — over a transcript.
The issue was later closed as completed, after official auto memory shipped. The current documentation still says those notes are not shared across machines or cloud environments. A teammate's agent does not automatically see what yours learned. Project context should hold what the code does not already say: decisions, constraints, and reasons. Copies of the repo into memory go stale.
Keep a current picture the next session can read
A practical split looks like this.
- Put the current picture where the next session will actually read it. Keep it with the project, in a form that can be reviewed, not only in one chat or on one machine.
- Write the decision, the constraint, and the why. Leave out what the next agent can read from the code or from git history. Those copies drift.
- Use a fresh session as a check. If it cannot recover the current answer without the original conversation, the record is not self-contained yet.
- When the truth changes, rewrite the current answer. Keep an append-only trail of how you got there.
MindMux captures background, decisions, and constraints from the conversations you already have with agents, and writes them into a Markdown project brain. Agents read that brain over MCP before they start a task. To connect Claude Code or Codex, install @mindmux/cli and run mmx setup in the project. mmx setup --local gives a local Brain without Cloud login. The open, local counterpart, brain.md, is a brain/ folder in your repo plus a small CLI (@mindmux/brain-md). Every page keeps a compiled_truth, the current answer, and an append-only timeline of how you got there. It is git-native and agent-agnostic: Claude Code, Codex, Cursor, and anything else that can read a file.
This is cross-session, cross-tool decision memory. If a decision made yesterday already shows up in the next session, on the next tool, it is working.