A memory only one agent can see is not project memory yet
Public issues show project knowledge sitting in one coding agent's local memory, invisible to the next tool. Official docs put those stores in the home directory. Durable decisions belong where every agent can read them.
Public developer issues and discussions often come back to a decision that was already written down, then vanished the moment someone opened a different coding agent. The note was real. It lived in one tool's local memory. The next tool never saw it.
The other agent cannot see the note
In a public issue on flick-drop, Felix-LeeSM put the problem in one sentence: project knowledge kept in one agent's local memory is invisible to other agents, specifically Codex and Gemini. The plan is to move durable facts into files in the repository, AGENTS.md and docs/, and to split documents that exceed what agents will load. The issue notes that Codex reads 32 KiB of AGENTS.md by default. OpenAI's Codex documentation says the same: Codex stops adding project instruction files once their combined size reaches project_doc_max_bytes, which is 32 KiB by default.
The size cap is a reason to keep a standing brief short. It is not the main failure. The failure is that the knowledge started in a store only one agent could read.
Each tool keeps its own store
The official docs make the split plain.
Claude Code writes auto memory under ~/.claude/projects/<project>/memory/ in the home directory. The first 200 lines of MEMORY.md, or the first 25KB, whichever comes first, load at the start of every conversation. Among the notes Claude may save is a project type: ongoing work, deadlines, and decisions it cannot derive from the code or from git history. The same page is direct about the boundary: auto memory is machine-local, and the files are not shared across machines or cloud environments.
Codex keeps a separate local memory store. ChatGPT web uses ChatGPT memory; local Codex clients use their own files under the Codex home directory, by default ~/.codex, in memories_v2/ or memories/. The docs tell you how to treat that store: keep required team guidance in AGENTS.md or checked-in documentation, and treat memories as a helpful recall layer, not as the only source for rules that must always apply.
Gemini CLI makes a similar cut. Shared project instructions go in repository GEMINI.md files. Private project notes go in a per-project private memory folder. Cross-project personal preferences go in the global ~/.gemini/GEMINI.md.
These pages describe each product's own memory. They are not a study of what happens when you switch tools. They still show the same shape: a decision written only into one agent's local store is available to that agent, on that machine. It is not yet a project record.
Copying the fact into every tool's file is not the same as sharing it
When the private store is invisible, a common workaround is to paste the same fact into each tool's instruction file.
In a public issue on synlynk, nikhilsoman reported that editing the Capability-Based Task Allocation table and its routing note currently requires hand-syncing across CLAUDE.md, AGENTS.md, GEMINI.md, and .cursorrules. The issue is open, and it is filed as tech debt. The cost is already visible: one change, four files.
That is not a shared memory. It is four copies you have to keep in step. A later edit that lands in only one of them leaves the next agent reading a different version.
Put the decision where the next tool can read it
A practical split looks like this.
- A preference that only matters in one tool can stay in that tool's local memory. How you like to see diffs in Claude does not need to travel to Gemini.
- Decisions, constraints, and project background that the next session or the next tool will need should live in the repository, in a place every agent can actually read. Official docs already point team guidance at checked-in documentation, not at a private recall layer.
- Do not treat one agent's private notes as the project record. If Codex or Gemini cannot see the decision, it is still a private note.
- If you are copying the same decision into four instruction files, it still does not have one home. Write it once. Point the tool briefs at that home. Do not restate it in each file.
MindMux itself makes a product in this space.
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. A note that lives in only one agent's home directory is useful to that agent. It is not yet the project's memory.