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.

[Get started free](https://console.context-mode.com/login) [Read the docs](https://context-mode.com/docs/memory)

Pick a moment

New session Fact changes New machine Old fix Fewer tokens

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 |

A store that only adds rows **23 of 24 stale** Context Mode **0 of 24 stale**

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.

Before: six top-ranked facts **262 tokens** Now: facts that fit **41 tokens**

Memory tokens per new conversation. Offline eval, 52 conversation openings. Facts sent: 5.96 to 1.23.

The problem

## Your agent starts each session empty. *You fill it again.*

**It forgets**

You told it the deploy target on Monday. On Tuesday, in a new session, it asks again.

**Two agents, two memories**

Claude Code and Codex keep separate local memories. A fact you give one is unknown to the other.

**Another machine, nothing to resume**

Built-in resume reads a file on this computer. On a new laptop or a remote box, that file is not there.

Proof

## The current value, not the old one. *In fewer memory tokens.*

**0 of 24** fact lists sent to the model held an old value, against 23 of 24 for a store that only adds rows

**262 → 41** memory tokens per new conversation, now that only facts that fit the turn go out

**8 of 8** questions answered right in new sessions with Memory on, without searching. Memory off: 8 of 8, with 9 searches. Claude Code with no gateway and no memory: 0 of 8, a baseline, not a test of its own memory.

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 →](https://context-mode.com/docs/memory#evaluation) [How we measured the tokens →](https://context-mode.com/blog/memory-with-two-clocks)

How it works

## One store on your account. *Every agent reads it.*

1. 01 **Work as usual** Your agent sends each request through the gateway. Context Mode keeps short facts, the conversation and the tool output.
2. 02 **A fact changes** The old row closes and stays as history. Only the current value goes out.
3. 03 **Any session asks** A new conversation gets the facts that fit the turn, in either agent, on any machine.

Both agents, on both machines, read the same store. The first request of a new conversation gets the current facts.

Compared

## 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 →](https://context-mode.com/docs/memory#compared) [Method and sources →](https://context-mode.com/docs/landscape)

Products used

## Memory, Session resume and Search. *One account.*

**[Memory](https://context-mode.com/memory)**

Keeps facts from your sessions. A changed fact closes the old one. [Memory docs](https://context-mode.com/docs/memory).

**[Session resume](https://context-mode.com/session-resume)**

One command rebuilds a conversation as a native session on any machine. [Docs](https://context-mode.com/docs/session-resume).

**[Search](https://context-mode.com/search)**

Finds last week's fix from another agent, so you do not explain it again. [Docs](https://context-mode.com/docs/search).

Related: [Instructions](https://context-mode.com/instructions) for rules you set once yourself · [Recall](https://context-mode.com/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](https://context-mode.com/pricing).

[Get started free →](https://console.context-mode.com/login) [Read the docs →](https://context-mode.com/docs/memory)