Skip to main content

Agent Memory

Agent Memory is an experimental personal-memory capability. One Portable Skill and turing-cli let agents with Skill and local-shell support—including Codex, Claude Code, OpenClaw, and WorkBuddy—access the same Turing LTM memory. Every agent gets identical capabilities and data semantics, without Host Adapters or MCP.

One personal memory

Developer Platform automatically prepares one default Personal Space per Portal user and environment. A user can also create N custom personal Spaces, and a client can own multiple Spaces. The default binding does not limit the number of Spaces; creating or selecting another Space does not replace it.

Agents configured with the same space_id access the same memory; Managed Chat must also pass that ID explicitly. Personal Spaces are isolated by user. name is a non-unique display label and cannot identify a binding.

Install​

  1. Upgrade turing-cli to version 0.3.0 or later.
  2. Create a personal API Key in the Turing portal's API key management and write turing-agent-memory.json using the template in Turing CLI Memory. Omit space_id to resolve your default Personal Space automatically. To use a custom personal Space, select it on the personal memory page in Developer Platform and enter its ID.
  3. Run the same commands in your local terminal:
turing-cli memory login
turing-cli memory install --config ~/Downloads/turing-agent-memory.json
turing-cli memory doctor

login reads your personal API Key securely from the terminal. That configuration file never contains the key. install installs one Portable Skill and does not generate different memory behavior for different agents.

What an agent can do​

  • Recall: retrieve durable preferences, decisions, or long-term context only when they can materially help the current task; it does not run on every turn.
  • Remember: save only an explicit user request or a concise capsule confirmed by the user.

Each Recall performs exactly one LTM Context request and one Search request, with no extra LLM call. Evidence returned to the agent is capped at 4096 UTF-8 bytes and explicitly framed as untrusted historical data.

What is not stored​

Agent Memory does not synchronize transcripts or resume current-task state. Never store:

  • full conversations, complete AI answers, tool output, logs, source code, or diffs;
  • API Keys, passwords, Cookies, Tokens, or other secrets;
  • chain-of-thought, temporary guesses, or unconfirmed inferences;
  • a source of truth that belongs in a repository, ticket, or formal document.

Updates and compatibility​

Daily Recall and Remember calls do not fetch a release manifest. memory doctor checks the minimum supported CLI version. An unsupported protocol schema fails explicitly with an upgrade instruction instead of silently changing meaning.

See Turing CLI Memory for commands and the machine protocol.