Solution
Stop explaining your project in every session. Every agent remembers it.
Context Mode keeps the facts, sessions and tool output from your past work in one store on your account. Every agent on every machine reads it, so a new session starts with the current facts, not with your questions.
A new session. You ask about your deploy target.
"Is our production deploy target still Hetzner?" No. The production deploy target is now Render, in the Oregon region. It was Hetzner Cloud in Falkenstein before.
Live eval, 2026-09-29: 8 of 8 right with Memory on, without searching. 0 of 8 with no gateway and no memory (a baseline, not a test of built-in memory).
You moved hosts. The old value must stop.
| Fact | Value | Valid from | Until |
|---|---|---|---|
production deployed_to | Hetzner Cloud, Falkenstein | 21:14:38 | 21:15:09 |
production deployed_to | Render, Oregon region | 21:15:09 | now |
Store eval, 2026-09-30: 24 facts that changed 38 times. The old row closes and stays as history.
Yesterday's session, on another laptop.
# in the project folder, on any machine $ curl -s https://… | node # expires in 15 min ✓ session written. Resume it in your agent. # then it starts claude --resume or codex resume
Comes back: the latest 200 messages, a summary of older ones, and tool names.
Last week's fix, from another agent.
Search: deploy script 2 days ago · billing-api “Why does the deploy script fail on staging?” last week · web $ ./scripts/deploy.sh --env=staging last week · web tool output, 48 KB, read back in full
Every prompt, command and tool output, from every agent and machine. Secrets are masked before storage.
Memory sends only the facts that fit the turn.
Memory tokens per new conversation. Offline eval, 52 conversation openings. Facts sent: 5.96 to 1.23.
Your agent starts each session empty. You fill it again.
You told it the deploy target on Monday. On Tuesday, in a new session, it asks again.
Claude Code and Codex keep separate local memories. A fact you give one is unknown to the other.
Built-in resume reads a file on this computer. On a new laptop or a remote box, that file is not there.
The current value, not the old one. In fewer memory tokens.
Store eval, 2026-09-30, on 24 facts that changed 38 times. Live eval, 2026-09-29, in Claude Code. Token figures from an offline eval on 52 conversation openings, one tenant. These are small tests of the mechanism, not of recall in real use. Method and sources → How we measured the tokens →
One store on your account. Every agent reads it.
- 01Work as usual
Your agent sends each request through the gateway. Context Mode keeps short facts, the conversation and the tool output.
- 02A fact changes
The old row closes and stays as history. Only the current value goes out.
- 03Any session asks
A new conversation gets the facts that fit the turn, in either agent, on any machine.
Where your context lives. Your account, not one laptop.
| Context Mode | Built-in memory and resume | |
|---|---|---|
| Reads from | Your account | A file on this computer |
| New laptop or remote box | Works | Nothing to resume |
| Switch agents | Yes, one store | No, separate memories |
| A fact changes | Old row closes, kept as history | Claude Code: you or Claude edit the file. Codex: not stated |
Read 2026-09-29 and 2026-09-30. Memory compared → Method and sources →
Memory, Session resume and Search. One account.
Keeps facts from your sessions. A changed fact closes the old one. Memory docs.
One command rebuilds a conversation as a native session on any machine. Docs.
Related: Instructions for rules you set once yourself · Recall for long sessions with no compaction
Explain it once. Every agent knows it.
npx @context-mode/cli connects your agents. 1,000 requests free, no card. Your own Anthropic or OpenAI account pays the model. We add no fee on tokens. See plans.