MCP surface trimmed to memory CRUD. The MCP/chat surface your tools connect
to is now six memory-CRUD tools —
save_context,
save_session, get_context,
list_context, resume_session, and
forget_context — pure memory CRUD plus session handoff
(save_context also updates an existing memory in place via a stable topic,
so there’s no separate update tool) — plus recall_memory, the one generative
tool, which rewrites a draft prompt using your saved context.The rest moved to the web app (the capabilities didn’t go away):
exporting your context is now an Account action in the dashboard, team
promotion is the Promote action in the Teams UI, and Library
ingestion/search/browse and data-source connections are managed on the
Library and Data Sources pages. Keeping the agent surface minimal means fewer
tool schemas crowding your tool’s context window.A tighter tool surface, and resume-where-you-left-off. The Library and data
source tools collapsed into a smaller, clearer set —
kb_ingest (text or URL),
kb_query (search, or fetch a page by slug), kb_browse (catalog or lint), and
one sources tool for connect/list/sync/disconnect — so
fewer tools crowd your agent’s context.New resume_session makes cross-tool handoff a single call:
end a session anywhere with save_session, then pick it up in another tool —
it lists your recent sessions and loads one in full.get_context gained a recency mode (favor the latest over
the most relevant), the destructive tools (forget_context, promote_to_team,
sources disconnect) take a confirm: false dry-run that previews without
acting, and list_context now shows cross-agent reuse — how often an entry has
been retrieved and the tool that last used it.Connect your tools as data sources. Connect an upstream MCP server
(Notion, Drive, Slack, or any custom server) and index its content on demand
into your Library — enriching the same context every tool draws on, not
relaying the source. Via
kb_connect_source / kb_sync_source.Memory types, including ephemeral tasks. Every entry now has a
memory_type — fact, event, instruction, or
task. Tasks auto-expire after a short TTL, so “what I’m working on now” never
hardens into stale knowledge. get_context can filter by type (e.g. just your
runbooks), and save_session classifies each item it distills.Two more tools. list_context to browse what’s stored and
forget_context to delete an entry for good — completing a
remember / recall / list / forget / export surface.Sharper hybrid retrieval. get_context now fuses four signals with weighted
RRF: semantic (with query-expansion at save), keyword (full-text search with
stemming), exact fact-key, and recency — your tool gets the most relevant
context first, no synthesis or extra round-trips.Cursor integration. Connect Cursor with a one-line installer; lifecycle
hooks auto-load and save context per session. See
Connect your tools.OpenAI Codex integration. Native remote MCP plus session hooks — Codex runs
the OAuth flow itself, no bridge required.Claude Code plugin. Bundles the MCP server, automatic get/save hooks, and
the
/onelamp command (aliased /ol).Passwordless sign-in. Sign in with an email one-time code, Google, or
GitHub.Closed alpha. The personal MCP endpoint goes live: link your AI client once
and it gets and saves context through save_context, get_context, and
export_context.