Agent Memory Product Model
Current implementation of profile-independent memory. Native compatibility was checked with Codex 0.153.4; model adherence remains dependent on the Agent.
The Hub owns long-term memory independently of Agent Profiles and Runtimes. memory.enabled is an experimental feature, disabled by default. The Feishu management experiments page controls this single Hub-wide switch; CLI users can set it with amux system config set memory.enabled true. Save and restart the Hub to apply it; remote Runtimes need no separate opt-in. Memory uses UTF-8 Markdown files under the Hub's AMUX_HOME/memory, with no memory database or persistent search index. Non-isolated installations use global/; when the Hub's user-isolation policy is enabled, the authenticated effective principal owns a separate users/ directory. Runtime availability and helper readiness never select a global fallback. Session access alone does not authorize access to another principal's memory. Service executions use their effective service principal.
The CLI and MCP tools search, read, conditionally write, forget and export documents through the Hub. Paths are relative to the server-derived scope and restricted to safe components: absolute paths, traversal, symlink escapes, hidden (dot-prefixed) or empty components, and over-long or non-UTF-8 names are rejected. Search matches literal keywords and returns bounded snippets; scanning limits are reported as incomplete results. Reads return bounded text, continuation information and a whole-file hash. Writes use that hash as a precondition, or absent for creation. A per-scope write lock and same-directory atomic replacement prevent lost concurrent updates. Multiple-file edits are not a transaction. Deleting a document does not erase already-read conversation history or duplicates in other documents.
Supported Agent launches receive one stable memory protocol and a frozen scope precondition. New native conversations read a small MEMORY.md through tools (the entry budget applies only to that exact filename); other material is read on demand. There is no per-turn automatic retrieval or prompt rewriting. The protocol is independent of profile projection. Existing unmanaged native conversations do not silently adopt a new memory policy. The managed Runtime keeps the scope fixed for the conversation; requests with an incompatible frozen scope fail. The binding is a non-secret client precondition, not an authorization token: every request is independently authorized by the Hub. Normal CLI users may manage their current authorized scope.
Managed Codex conversations disable native automatic memory use and generation with launch-local controls. Claude disables automatic memory, including native subagent persistent memory, through its supported environment switch. The reviewed Pi core has no built-in automatic long-term memory; custom memory extensions require separate review. Other providers need a verified instruction and native-memory contract before adopting amux memory. User global configurations and native memory files are not directly edited or migrated. Isolated native home paths must not inherit shared CODEX_HOME, CLAUDE_CONFIG_DIR or PI_CODING_AGENT_DIR overrides. Provider input usage already includes returned memory; Feishu titles do not add a memory counter.
Memory requires a runnable amux CLI on the Runtime. Linux DAC launches copy the CLI and minimal Hub connection config into the current run's agent-config/memory/ through the existing target-UID materializer; shell and MCP clients use that copy. The ordinary run cleanup reclaims it. Other launches use the installed CLI. Missing tools fail visibly before starting the Agent.
Disabling the experiment rejects all memory data RPCs and new bindings while retaining stored files. Existing conversations keep their frozen protocol and native automatic-memory suppression, including on resume: disabling access must not make their history eligible for native extraction. They cannot read or write amux memory while the Hub switch is off. MCP discovery checks Hub capability once at startup; every data call still enforces Hub policy. Re-enabling permits existing compatible bindings and new conversations; legacy unbound conversations need a new conversation. Saving settings without restarting leaves the running policy unchanged.
If tools report memory unavailable, the fixed protocol directs the Agent to continue without recall or repeated retries, and never fall back to another automatic-memory store. Existing bound launches retain the CLI dependency even while access is disabled. Cached MCP memory continuations also check the Hub switch before returning another page.