The moat gets deeper, newest first. Most of these are memory-trust features: the layer between “my AI remembered something” and “my AI remembered the right, current thing.”
One 0..1 number per note, folded from signals we already track: anchored to real code lifts it, stale or aging sinks it, and notes your agents actually reuse earn trust. Recall, list and MCP all surface it, so a caller knows what to lean on and what to double-check.
A note can now carry an explicit validFrom / validTo, so “this was true in v1.2” stays queryable and out-of-window memory is down-weighted in recall instead of deleted. Hard windows complement the soft temporal decay already in place.
Within the same scope, notes that are about the same thing but disagree are surfaced as review candidates — by embedding distance alone, no model required. Connect a completer and it upgrades to a real contradiction check that proposes which note supersedes which.
After a sync, SecondOS reads the map diff and recent decisions and proposes memories you accept or reject — assisted, never auto-written, and map-only so it never touches your source. Proposals need a connected model; without one it degrades quietly to no proposals.
A navigable view of how notes attach to real symbols, tinted by trust and validity so stale, low-confidence memory is visible at a glance. The memory finally has a map of its own.
A workspace-scoped memory that isn't bound to one project — shared conventions that fold into recall across every repo in the workspace. One home for the rules your whole team keeps re-explaining, project after project.
Invite people into a project with roles; they read the same structure-aware map and memory you do. Sharing covers the map and the notes — never your source, which still never leaves the machine that synced it.
Pull an exact symbol into a question with @, and answers cite the precise source lines they leaned on, numbered. Grounded in the map, and traceable straight back to where it came from.
The MCP surface grew memory-hygiene tools — history, restore, reconcile — alongside remember and recall, plus a single generic run_action_tool that fronts every connected integration behind the same preview → confirm gate.
Ask a question or request a file summary in the app and it runs on Qwen3 235B, built entirely from the map — signatures and doc comments, never raw bodies. No model connected? It falls back to the offline structural answer.
Connect ClickUp with an API key (hashed and encrypted at rest). Read tools are safe by default; every write drafts a preview you explicitly confirm before it runs, with a full audit log. Exposed over both chat and MCP.
A push webhook re-extracts only the changed files on the runner and updates the cloud map and vectors. Sync once, stay fresh forever — and the source-never-leaves invariant holds on the runner too. Auto-syncs count toward your quota.
A new pass in the drift pipeline spots contradicting notes on the same scope, supersedes the newest and marks the stale one down. Your recall stays clean as memory grows, without heavy solver machinery.
Walk the supersede chain for any note with GET /memories/:id/history, diff two versions, and restore an earlier one — as a new supersede, so nothing is ever deleted. Memory accretes instead of drifting.
When a note names a scope ref, we resolve it against the latest sync graph. Real symbol → verified. Unknown ref → saved as unverified with a clear warning, never silently dropped. This is what makes memory bound to your code.
Retrieval now weights notes by age. Fresh advice outranks a six-month-old note at the same cosine score, and stale, scope-broken memory sinks to the bottom instead of resurfacing.
Summaries and /ask answers now run on a single stronger model (Qwen3 235B on Bedrock). Prompts are built from the map — signatures and doc comments — never raw source bodies.
> daemon: değiştim. çoğu şey iyiye. birkaçı sadece değişiklik olsun diye değil.