What changed
Claude Code and the Gemini API both now speak remote MCP natively (FACT: Gemini managed-agents announcement cites remote MCP support; Anthropic ecosystem tooling is MCP-native), so a team memory layer can be shipped once as a protocol endpoint instead of per-tool plugins. Simultaneously a concrete, well-specified pain surfaced on HN: cofounders' Claude Code agents are drifting apart with context trapped per-machine (FACT: Ask HN post).
Why now
Agent context is newly valuable and newly siloed: individual developers have months of accumulated agent memory with zero team-sync story. Remote MCP just became the universal socket, and provider-hosted agent orchestration (Gemini managed agents) signals the whole market is moving to persistent, long-running agents that need durable shared memory. Window exists BEFORE Anthropic ships native team memory (HYPOTHESIS: they likely will — this is the core kill risk).
Converging signals
(1) Anthropic-compatible tooling is now backend-agnostic, including local models via Ollama — teams standardize on the Claude Code interface regardless of model (FACT: ollama.com/blog/claude). (2) Gemini API adds background tasks + remote MCP, normalizing remote MCP as the tool/memory transport (FACT: blog.google announcement, headline-level). (3) Direct PAIN: two cofounders maintaining drifting 'company brains' with only a Google Drive folder as workaround (FACT: HN item 48848840, similarity 0.783).
Customer pain
Teams of 2-10 running per-developer coding agents each accumulate strategy, customer notes, decisions and conventions locally. The context is trapped per machine, duplicated work is invisible, memories drift and conflict, there is no history/merge, and no shared-vs-private boundary. Current workaround is a synced folder of markdown, which has no merge semantics, no provenance, and leaks private context (FACT for one team from the HN post; breadth of pain is HYPOTHESIS).
Who pays
AI-native startup founders and small dev teams (2-20 seats) already paying $100-200/seat/month for Claude Max/Team — they demonstrably spend on agent tooling and feel this pain daily. Buyer = the technical founder/lead, reachable directly on HN/X/Discord without procurement (HYPOTHESIS: willingness to pay for the memory layer specifically is unproven; no HIRING/SPEND evidence was provided).
Solved today
Shared Google Drive/git folder of CLAUDE.md and markdown notes; pasting context between machines; each dev re-deriving answers; early OSS memory servers (Mem0/OpenMemory, Zep, Letta, Cognee) that are mostly single-user or generic RAG memory rather than team-shared/private-split with merge semantics (competitor characterization is HYPOTHESIS pending verification).
Why current solutions are bad
Folder sync has no merge/conflict resolution, no history, no provenance ('who decided this and when'), no shared/private boundary, and no MCP-native retrieval — the agent can't query it semantically. Generic memory servers are per-agent, not team-scoped with access control.
Proposed product
A remote MCP server (hosted, plus self-host option for trust-sensitive teams): namespaced team memory with (a) explicit promote-to-shared gesture from local context, (b) append-only decision log with provenance and history, (c) merge/conflict semantics for divergent facts, (d) semantic retrieval tools exposed via MCP, (e) private-by-default local memory untouched. Works identically for Claude Code, Gemini agents, and anything MCP-speaking.
MVP version
Single-tenant hosted MCP endpoint: Postgres + pgvector, three MCP tools (remember_shared, recall, log_decision), API-key auth per team, a plain web UI showing the shared log with diff/history, and a CLI to promote local CLAUDE.md sections to shared. No merge UI yet — last-write-wins with full history. ~3-4 weeks of AI-assisted solo build; founder has shipped exactly this stack shape before.
30-day build
Ship MVP; answer the HN thread and DM the OP; recruit 10 design-partner teams from HN/X/Claude Code Discord with free 60-day pilots in exchange for weekly feedback; instrument retrieval usage to prove the agent actually pulls shared memory in real sessions.
60-day build
Add shared/private namespaces, decision provenance, and conflict surfacing (show both versions + ask-the-team). Publish to MCP server directories/registries. Write the 'stop maintaining two company brains' post. Convert pilots to paid at $49-99/team/month; add self-host license ($499/yr) for teams that refuse hosted memory.
90-day revenue plan
Target 15-30 paying teams = $750-3,000 MRR by day 90-120. Honest framing: this is a niche-wedge revenue ramp, not a fast-cash play; the bet is riding Claude Code's growth curve. If <5 teams convert from ~10 engaged pilots, kill or open-source it for reputation and move on.
Distribution path
Direct: reply in the originating HN thread, launch post on HN/X, MCP server directories (official registry, mcp.so, Smithery), Claude Code/Cursor Discords, SEO on 'share Claude Code context with team'. Demonstrated-value motion (live demo of two agents sharing a decision) matches founder's selling style. No enterprise channel needed.
Pricing hypothesis
$49/team/month up to 5 seats, $15/additional seat; $499/yr self-host license; free single-user tier as top-of-funnel. Anchor against the $100+/seat teams already pay for the agents themselves.
Technical difficulty
Moderate-low for MVP (FastAPI/MCP SDK + Postgres/pgvector — founder's existing stack). Real difficulty is product semantics: what auto-promotes to shared, conflict resolution UX, and preventing memory-poisoning garbage-in. 6/10 overall.
Legal / regulatory risk
Low regulatory risk, but the product stores customers' strategy and customer data — needs solid tenancy isolation, encryption, DPA boilerplate; the self-host option is the pressure valve. One prompt-injection/memory-poisoning incident could kill trust (HYPOTHESIS on likelihood).
Platform dependency
High and this is the #1 risk: Anthropic already has memory primitives and team plans; a native 'shared project memory' feature would gut the standalone value overnight. Mitigation is multi-agent neutrality (Claude + Gemini + local models via MCP) and the self-host/privacy wedge, but the dependency is real.
Founder fit
Mixed. Fits: micro-SaaS/API shape, fast AI-assisted prototyping, demonstrated-value sales, founder is himself a heavy Claude Code user so has authentic domain credibility. Doesn't fit: this is NOT his proven government-portal forced-buyer pattern (lesson, conf 0.80) — no mandate, no deadline, discretionary dev-tool spend, and dev-tools is a crowded, fashion-driven market where distribution favors people with existing dev audiences. 5-6/10.
Breakout potential
If MCP-native team memory becomes the standard pattern before platforms internalize it, this becomes the 'shared brain' layer for agent-using companies — genuine expansion into org knowledge-ops. But the same dynamic makes it a likely acquisition/fast-follow target for Mem0/Zep/Anthropic. Moderate breakout, high fragility.
Final recommendation
CONDITIONAL BUILD — validate before committing. The pain is real and articulately specified, the MCP timing is genuinely good, and the MVP is squarely inside the founder's build capability and budget. But with a single PAIN datapoint, no spend evidence, a crowded memory-server field, and high platform-absorption risk, this is a B-/C+ opportunity: spend ≤2 weeks on a validation sprint (HN thread engagement + 15 founder interviews + landing page with pricing) before the 4-week MVP. Do not treat as the flagship bet; it does not match the founder's proven forced-buyer pattern.
Next action
Today: reply to HN item 48848840 and DM 10 teams from that thread and Claude Code Discord offering a concierge pilot; stand up a landing page with $49/mo pricing to measure signup intent before writing product code.