The session is short. The project runs for months. Here's how to keep the thread

The issue, as reported

It is one of the most widely reported problems on the Claude Code, Codex and OpenCode issue trackers: in a sweep of public issues, several hundred developers describe a long-running project outgrowing the session, many of them since July.

The pattern is the same across tools. On day one the agent works out the architecture and the decisions with you. On day two the session is gone and all of it has to be explained again. Some developers say they have stopped attempting multi-day tasks for that reason. People with hundreds of sessions behind them say they can no longer find the one where they stated a requirement. Heavy users count dozens of compactions a month, and several built their own memory system.

Running agents in parallel makes it worse. Close to a hundred developers describe several sessions or sub-agents working on one project with no way to see each other's findings: one session changes a schema, and another carries on building against the old one. A sub-agent given a narrow role, such as testing, has been reported drifting out of that role after an automatic compaction.

Why it happens

Every session is a fresh start. Anthropic's documentation says so directly: each Claude Code session begins with a fresh context window. A new session loads your CLAUDE.md or AGENTS.md and a capped memory index. It does not read your earlier conversations, where the reasons behind the decisions live.

Three things make that worse over months:

• The old conversations age out. Claude Code keeps transcripts on disk, but deletes them after 30 days by default (the cleanupPeriodDays setting). On a six-month project, most of the history is gone unless you change it.
• Long sessions are summarised. Compaction replaces the conversation with a summary, and when you resume a long, idle session from a summary, the docs note that whatever the summary leaves out is no longer in context.
• Sub-agents start blank and report back a summary. In Claude Code, each sub-agent starts with a fresh, isolated context window and only its final result returns; sub-agents compact the same way the main conversation does. Codex's guidance is to return summaries from sub-agents instead of raw intermediate output. What a sub-agent learned on the way is kept in its own transcript, which nothing reads again.

How to fix it today

These steps use only the tools themselves and follow their current documentation.

1. Keep the transcripts for the life of the project

   In Claude Code, raise the retention in ~/.claude/settings.json. A project's own settings file can override yours; our guide to the 30-day deletion covers the checks.

   {
     "cleanupPeriodDays": 3650
   }

2. Name sessions by workstream

   Start with claude -n billing-migration, or run /rename billing-migration in a session, then come back with claude --resume billing-migration. In the picker, type to search and press Ctrl+A to see every project. In Codex, codex resume reopens a recent chat.

3. Keep a decision log in the repository

   One dated line per decision, with the reason and what it replaces, committed with the code. Add a line to AGENTS.md telling both tools to read it before changing anything it covers.

   2026-08-04  Postgres, not SQLite: reporting needs joins
   2026-09-12  Retries in the client, not the proxy (replaces 08-20)
   2026-09-29  Invoices are immutable once sent

4. Hand over at the end of each session

   Keep a short handoff file (now, next, what failed) and load it automatically at start-up. Our article on re-explaining your project has the file and a working SessionStart hook.

5. Give sub-agents somewhere to write

   A Claude Code sub-agent (.claude/agents/<name>.md) with memory: project keeps notes in .claude/agent-memory/<name>/, and the first 200 lines of its MEMORY.md load next time. For parallel sessions and Codex sub-agents, add one instruction to AGENTS.md: append findings and decisions to docs/findings.md before stopping.

   ---
   name: test-runner
   description: Runs the test suite and reports failures
   memory: project
   ---

6. Let live Claude Code sessions talk

   From v2.1.224, cross-session messaging is on with nothing to enable. Run /list-agents to see your other sessions, then ask: "Tell the session working on the payments API that the schema migration finished."

What the fix doesn't cover

How Axiom Lift solves it

Axiom Lift keeps one history of the project, built from the sessions your tools already save, so the record doesn't depend on anyone's discipline.

On a project that runs for months, the plan matters. On the free plan, your AI can reach the last six months of history (up to 1 GB). Pro keeps all history and tells you when a later conversation changed an earlier decision. Automated agent runs are told apart from the sessions you type, kept and searchable, and out of the way.

For parallel agents, Axiom helps only in part. What one session settled is there for any later session to ask, in any of the three tools. But it is not a live channel between agents running at the same moment, and it is not a scratchpad sub-agents write to mid-task. Use the tool's own messaging for live coordination. Axiom also doesn't replace AGENTS.md for rules that must always apply, and it doesn't cover Cursor, or ChatGPT and Claude in a browser.

Install is one command in a desktop terminal (Mac; Linux and Windows in beta; needs Python 3). It shows you what it found before anything is sent, it's free, and axiom uninstall removes it.

Claude, Codex and OpenCode are trademarks of their owners; Axiom works alongside them and is not affiliated.