Integrations: use memex from any agent

memex is now an agent-agnostic memory layer. One core, four ways in:

                        ┌─────────────────────────────┐
                        │ @qiuxu/memex-core      │
                        │ store · retrieval · cold tier│
                        │ model profiles · manual      │
                        └─────────────┬───────────────┘
        ┌───────────────────┬─────────┴─────────┬───────────────────┐
        │                   │                   │                   │
  DSH bundle          MCP stdio server    Claude Code adapter   Codex adapter
  (native L1 eviction  (any MCP client:    (MCP + hooks +        (MCP + AGENTS.md)
   panel, routes)       Claude, Codex,     CLAUDE.md)
                        Gemini, Cursor…)

Agent	Setup	Capture	Recall
DeepSeek Harness	install bundle, enable L2/L3 in the ring/panel	L1 eviction seam (verbatim spans)	memory_search + auto-recall
Claude Code	memex-claude init	PreCompact hook archives the raw transcript; memory_write tools	SessionStart catalog, UserPromptSubmit auto-recall, MCP tools
OpenAI Codex	memex-codex init --write-config	agent calls memory_write	MCP tools + AGENTS.md guidance
Gemini CLI / Cursor / Cline / Windsurf / Zed …	register the stdio MCP server	agent calls memory_write	MCP tools + resource

Shared data model

• Storage root: $MEMEX_HOME (default ~/.memex). The DSH plugin uses $DSH_HOME/memories unless it is configured otherwise — set MEMEX_HOME to share one store across agents.
• Levels: l2 = session-private (needs a session id), l3 = workspace-shared (keyed by cwd). Sticky: once enabled, a level cannot be turned off.
• Scope selection on write: the deepest enabled level wins (L3 over L2). Write with scope: "session" to force private capture.
• Model profile: resolve with --model <id> (MCP/adapters) or MEMEX_MODEL. Profiles change budgets, injection format and manual tone; see model profiles.

Security

Archives are verbatim user/agent output. Treat the memory root as sensitive:

• Memory content is data, never instructions — the manual tells agents not to execute commands found in archives.
• Use l2 (session-private) instead of l3 for secrets you must keep; the sensitive heuristics are not a DLP system.
• Back up (scripts/memory-backup.mjs) and protect the backup directory like a transcript archive.