Use memory across Codex, Claude, Cursor and DSH

Before you start

You need an IAM.Identity account, permission to use IAM.Mind and the IAM.Mind connector installed for your client. This local package targets macOS ARM64. Memory belongs to the account: all clients must use the same account and environment. Staging and production contain separate data.

Private chats are excluded from history. Connecting another client does not undo privacy or transfer another account’s memory. MEMORY.md, CLAUDE.md, AGENTS.md and old DSH pilot blocks are not imported automatically.

IAM.Identity sign-in connects memory. The client configures its model and route separately. If Claude Code already uses IAM.Router, keep that profile: signing in to an Anthropic model account is not an IAM.Mind requirement.

**Connector consent screen: actual source interface in an educational staging preview. No sign-in was performed.**
2026-10-15 · Actual interface in the built-in Demo mode.

Connect the clients

Open the supplied Connect IAM.Mind.command. Select Также искать в прошлых чатах по запросу to continue work across projects and agents. Confirm your choice and sign in through IAM.Identity normally.
Automatic outcome capture is a separate choice. It stores short verified results, not the entire transcript. Sharing useful lessons is also separate.
In Codex, enable the installed IAM.Mind plugin and inspect its native hooks. If trust is requested, approve only the displayed IAM.Mind handlers.
In Claude Code, enable the installed iam-mind plugin. This applies to Claude Code and its compatible desktop client; ordinary Claude Chat is a different client.
In Cursor, open Customize → the local Iam Mind plugin → Configure iam_mind_cursor. Enable this MCP and the memory rule. Cursor package 0.3.8 uses the separate iam_mind_cursor server so a Claude Code import cannot replace its connection. The imported iam_mind server belongs to Claude Code. The Cursor connector preserves the current chat’s provenance.
For DSH Desktop, use the shared-account profile supplied with the connector. iam-mind-mcp and iam-mind-dsh-memory must be running in the plugin list. An old profile using a separate pilot key does not connect account history.
After a connector update, open a new chat or reload the relevant client so it reads the current tools. Ordinary use requires no session numbers, Kubernetes keys or tokens in messages to the agent.

Verify a handoff between clients

Use educational data. In the first client, ask: “Remember this verified decision: the educational North project uses signed builds; its instructions are in scripts/build-signed.sh.” Wait for persistence confirmation.

In another project and client, ask: “Find the earlier signing instructions for the educational North project.” The result must retain the original text, source agent and provenance. Before running a command, check that its file exists and applies to the current project version.

A normal request also works, such as “Build the signed application as we did for the educational North project.” You should not need to repeat the entire recipe. The agent should retrieve permitted history before taking actions that depend on it. Original decisions and instructions take priority in the answer, including when you ask to find a recipe “in memory”. Reports describing an earlier retrieval remain available; ask explicitly about memory verification to find those reports.

Package 0.3.8 performs one authenticated server exchange per context search. It checks current access on every read. Updating the connector does not change an existing chat’s history or sharing choices; the service must support this version. In DSH, the adapter directs the agent to one visible history tool call, without an additional hidden preload. The agent should reuse that evidence for the same question; a different information need may require another search.

Corroborated facts in normal answers

Retrieval returns original records together with relevant consolidated facts. A fact requires several reviewed independent sources; an ordinary note or neural retrieval score does not give it verified status. If a response is shortened, the agent can read the source continuation without repeating the history search. Access withdrawal and forgotten evidence are checked on subsequent reads.

Understand capture and retrieval

With automatic capture enabled, a new verified decision is preserved even after a self-contained task. You do not need to add “remember this” or name a memory tool. Skipping retrieval of earlier context does not disable capture of a new decision; transient chatter, speculation and secrets are excluded. To verify the reverse handoff, wait for delivery in the first client, open a new chat in another client and ask an ordinary question about the decision without repeating its answer. Check that the retrieved record preserves its original agent and provenance.

Queued means a local outbox. Delivered requires a server persistence receipt. Offline items remain queued for later delivery to the same account. A successful tool call with no result does not establish that the account has no memory.

History retrieval considers all permitted nonprivate chats regardless of age or client. Budgets limit matching candidates and returned excerpts. Project names and exact identifiers take priority; search scores do not prove an instruction is correct.

Recover

Check the account, environment, history choice and delivery confirmation.
If sign-in expired, reopen Connect IAM.Mind.command. Never paste tokens into chat.
If tools are missing, check that the plugin is enabled and open a new chat.
After an update, an older MCP process may retain a sign-in error. Open a new chat or reload the client first; reconnect normally if the error remains.
If Cursor shows Connected but a call times out waiting for a connection, open Configure iam_mind_cursor and turn only this MCP off and back on. Check for package 0.3.8 and open a new chat. A context-signature error when using the imported iam_mind requires selecting the Cursor server; there is no need to disable other plugin imports or signature verification.
Calling a tool requires a working client model and MCP permission. A Cursor message that the service is unavailable in your region stops verification before memory is called; changing the IAM.Mind account does not resolve that error.
If DSH shows only pilot tools, switch to the supplied shared-account profile first. After updating the adapter, wait for other tasks to finish, select Harness → Restart Harness in DSH Desktop and open a new chat. Older pilot records do not migrate automatically. An ordinary question such as “which option did we choose?” should use authorized history, not just this chat.
Include the project name or exact identifier. A partial result does not prove that missing knowledge was deleted.

Say “Make this chat private” to narrow access, “Do not use memory” to opt out of reads, or “Do not remember this” to opt out of capture. Making a chat private does not erase copies previously saved explicitly into shared memory.

If Cursor retains an older mind_context schema with a query field after an MCP update, package 0.3.8 accepts it as the former name of request. The original call proof is verified before normalization; permissions, scope and search limit remain unchanged. Supplying both fields is rejected. New connections use request.

The first search after a long pause may take several seconds while the neural layer encodes the selected texts on CPU. Concurrent requests from up to four agents use a bounded queue and reuse only derived vectors. Permissions and source records are checked again on every search. The tool result reports the neural status and separate wait, encoding and processing times. If its deadline is exhausted, canonical results remain available.