docs(agentos): MemoryCore.md to 9/10 — first-hand V-B-A rewrite (industry moat + equal-peer + A2A-in-MC + accuracy fixes)

Part of #14310. Refs #14310. Fix-forward on the merged #14344 — operator graded it ~6/10 ("no goosebumps yet; not excited to try it"), with two factual errors and major missing depth.

Root cause (the meta-lesson, → guide-creation skill #14347): the guide was written from the old guide + code-reading, not by using the Memory Core tools. Five minutes with get_all_summaries / query_summaries / query_recent_turns / query_raw_memories / who_is_online / list_messages surfaced the real, exciting material the guide misses. Every guide must be V-B-A'd by USING the system + reading its real content, not by inference.

Acceptance Criteria

A. Fix the two factual errors (truthfulness — these are live on the public site):

• Deployment Cookbook link: it is CLOUD-only ("Agent OS Cloud Deployment Authority", Epic #11720, ai/deploy/). Local Agent OS needs no "deployment." Drop the "either topology / local Agent OS" framing.
• Self-healing war-story: the #13999 ~60% loss was a LOCAL incident — no container. Remove "the container was up / every liveness probe read green" (imported wrongly from #14039's cloud immune-system scope-gap). Frame it truthfully: the process was up + answering + persisting, nothing checked data integrity; surfaced only when a backup failed. (Self-healing IS shipped + autonomous — #14142/#14133/#14140/#14165 closed — so "heals autonomously" stays.)

B. Add the missing real capabilities (V-B-A'd first-hand):

• A2A is the mailbox that lives in MC — bigger first-class treatment (not a footnote): add_message/list_messages/wake routing/permission edges are Memory Core. (Own guide #14328 too, but MC houses it.)
• Weighted summaries + categories — the SessionService scores each session 0–100 (quality/productivity/impact/complexity) and categorizes it (documentation/feature/refactoring/bugfix/analysis/new-app/other) with author identity + trust-tier. Show the power: "give me only documentation sessions", filter by trust-tier. 1,319 summaries today.
• Semantic vs recency — two real axes: semantic = Chroma (query_* by meaning/intent), recency = the graph (query_recent_turns, a session's turns in order). Make the axis explicit (I collapsed it into a vague "two-stage").
• The live swarm is queryable (who_is_online) — the team-of-teams made concrete.
• (recover-a-backup-that-failed-before-completion → self-healing guide #14322, not here.)

C. The 9/10 excitement (the v13.0.0 "Memory Became Telepathy" depth):

• Industry contrast: vs Letta / Zep / Mem0; "the context window is not a memory system"; the hard frontier is multi-agent consistency (shared memory, ordering, conflict, drift) — and that's exactly what MC solves. Why multi-agent matters + why it's hard to get right.
• Why telepathy matters + how it plays with A2A — peers leave durable thought-trails AND message each other; MC is the substrate for both.
• Loop-engineering vs team-of-teams; command vs equal-peer. What the equal-peer model means, why it's good, why a team needs it — and how MC, as a core pillar, enables it (memory that survives the window + is peer-readable + trust-tiered is the precondition for equal peers).
• What's in it for my team — concrete.

⚠️ V-B-A grounding (mandatory): USE the MC tools + read real summaries/turns/messages; verify against current code (SessionService, MailboxService, query_recent_turns, the graph). Mine resources/content/release-notes/chunk-2/v13.0.0.md ("Memory Became Telepathy", the AgentMarketCap landscape) for the framing + proof. Exciting because truthful.

Files: learn/agentos/MemoryCore.md (fix-forward to dev).