What changed
Teams have gone from one person using Claude Code to multiple cofounders/employees each running their own local agent, and the accumulated context (strategy, decisions, customer notes) is now trapped per-machine and drifting (FACT: HN pain post). Meanwhile MCP has become the standard way to attach external stores to these agents, and providers (Gemini managed agents, Anthropic-compatible local stacks) now support remote MCP natively (FACT: Google blog headline; Ollama blog).
Why now
Claude Code adoption inside 2-10 person teams is recent enough that no default team-memory workflow exists; the workarounds cited are a Google Drive folder (FACT from the HN post). MCP support in both local and hosted agents means one memory server can serve every surface — a window that likely closes when Anthropic ships native team memory (HYPOTHESIS).
Converging signals
(1) Concrete team pain: two cofounders with drifting per-machine agent context, no shared/private boundary, no history (HN #48848840). (2) MCP as a universal attachment point, including provider-hosted remote MCP (Google managed-agents announcement). (3) Anthropic-API-compatible local backends broadening the Claude-Code-style user base (Ollama blog). Signals converge on 'agent memory is a team resource, tooling is single-player'.
Customer pain
Each teammate's agent starts from zero on questions a colleague's agent already solved; two 'company brains' drift apart; no way to share a subset while keeping personal/private context separate; no merge or history semantics. This is stated verbatim in the source, not inferred.
Who pays
AI-native startup teams of 2-10 who live in Claude Code daily. HYPOTHESIS: the buyer is the technical founder, paying $15-30/seat/month like they pay for Linear or Raycast. No spend evidence in input — no HIRING/SPEND items were provided, so willingness-to-pay is unproven.
Solved today
Checked-in CLAUDE.md files in the repo, ad-hoc markdown notes in a shared Google Drive folder (FACT: HN post), copy-pasting context between sessions, or open-source memory layers (Mem0/OpenMemory MCP, Zep, Letta) self-hosted by the more sophisticated.
Why current solutions are bad
Repo-checked files handle code conventions but not private-vs-shared boundaries, cross-project strategy/customer context, merge conflicts between two agents writing memory, or history/rollback. Drive folders have no MCP interface, no structure, no conflict handling.
Proposed product
A hosted sync service + open-source MCP server: each teammate's Claude Code connects to a local/remote MCP memory endpoint; memories are namespaced private/team/project; writes go through merge logic with full history; a small web UI shows/edits the shared brain; optional remote-MCP endpoint so hosted agents read the same store.
MVP version
Single-tenant hosted Postgres+pgvector store, an MCP server exposing remember/recall/search tools with private|shared scope flags, last-write-wins plus append-only history (no fancy merge), magic-link team invites, and a read-only web viewer. 3-4 weeks solo with AI assistance.
30-day build
Build MVP; recruit 10-20 design partners directly from the HN thread and Claude Code Discord/subreddit; free during beta in exchange for weekly calls; instrument which memory ops actually get used.
60-day build
Add the private/shared boundary polish, history/rollback UI, and team onboarding; publish the MCP server open-source as the distribution wedge; Show HN launch; list in MCP directories/registries.
90-day revenue plan
Convert design partners to $20/seat/month (teams of 3-5 → $60-100/mo each); target 20-40 paying teams ≈ $1.5-3k MRR by day 120-180. HYPOTHESIS — no pricing evidence exists in the input.
Distribution path
Reply-in-thread on the exact HN post, Show HN, MCP server directories, Claude Code community Discord/Reddit, open-source server with hosted-sync upsell. Founder sells through demonstrated value, which fits dev-tool distribution.
Pricing hypothesis
Free solo tier (local only), $20/seat/mo team tier (hosted sync, history, boundaries). Open-core: MCP server free, sync/history/team features paid.
Technical difficulty
Moderate and squarely solo-buildable: MCP server, auth, vector+keyword retrieval, append-only event log. The hard part is memory-quality UX (what gets remembered, staleness), not infrastructure.
Legal / regulatory risk
Low regulatory risk, but the product stores customers' strategy and customer notes — a breach is existential for trust; needs encryption-at-rest and clear data terms from day one.
Platform dependency
HIGH. The product's reason to exist is a gap in Claude Code; Anthropic already ships CLAUDE.md/memory primitives and is the single most likely party to ship native team memory, which would gut the standalone product overnight. MCP portability (works with Gemini managed agents, local models per the Ollama signal) is the only hedge.
Founder fit
Mediocre. This is a developer-tools infrastructure play sold to AI-native engineers in a fast-moving, crowded niche — far from the founder's proven edge (government-portal filing mandates, forced buyers, industrial/compliance niches). He can build it, but he has no audience advantage among Claude Code power users, and the accumulated-lessons prior says his highest-fit shape is regulation-compelled filing tools; this matches none of that shape.
Breakout potential
If agent-memory-as-a-team-resource becomes a category, the open-source MCP server could become a default and the hosted sync a real business — but that same attractiveness is why Mem0/Zep/Letta and Anthropic itself will contest it.
Final recommendation
PASS for this founder / WATCH the category. The pain is real and well-specified but rests on one data point, willingness to pay is unproven, the competitive gap is thin against free open-source memory layers, platform risk from Anthropic is severe, and founder fit is weak versus his proven government-portal forced-buyer edge. Only revisit if (a) multiple independent PAIN/SPEND signals accumulate and (b) Anthropic signals it will NOT ship team memory natively.
Next action
Spend one hour, not one sprint: reply to HN item 48848840 and message 5 similar teams asking what they'd pay monthly for hosted shared agent memory; simultaneously check Anthropic's changelog/roadmap chatter for native team-memory plans. If fewer than 3 of 5 name a price ≥ $15/seat, archive.