What changed
Claude Code adoption has spread to multi-person teams, but each member's agent accumulates its own local context (strategy, customer notes, decisions) trapped per machine [FACT: HN post 48848840]. Simultaneously, Anthropic-compatible local stacks (e.g. Ollama's Anthropic API compatibility) make Claude Code the de facto agent shell across backends [FACT: ollama.com/blog/claude], enlarging the installed base that shares this memory problem.
Why now
Two concrete pains surfaced within days of each other: teams whose agent 'company brains' are drifting apart with only Google-Drive-folder hacks [FACT: HN 48848840], and users who feel loss of control over agents and want audit/oversight tooling [FACT: HN 48846739]. No incumbent product cleanly addresses shared/private-partitioned agent memory with history; HYPOTHESIS: the window exists until Anthropic ships native team memory.
Converging signals
(1) Claude Code usable against multiple backends including free local models, expanding the user base [ollama.com/blog/claude]; (2) explicit team-memory pain: sync conflicts, drift, no shared/private boundary, no history [HN 48848840]; (3) explicit oversight/auditability pain for agents [HN 48846739].
Customer pain
Cofounders each maintain a divergent agent context; when one asks his agent something the other already worked out, it starts from zero — 'two company brains drifting apart' with only hacky file-share workarounds [FACT: quoted in HN 48848840]. Separately, operators can't see or replay what context/decisions their agents used [HN 48846739].
Who pays
HYPOTHESIS: 2–20 person AI-native startups and agencies where multiple people run Claude Code daily and agent context is business-critical (strategy, customer notes, conventions). Buyer is the technical founder; per-seat SaaS pricing. No direct evidence of payment yet — demand_evidence contains PAIN only, no HIRING/SPEND and no FORCED BUYER, so willingness-to-pay is unproven.
Solved today
Shared Google Drive folders of markdown, hand-maintained CLAUDE.md files in a git repo, copy-pasting context between machines [FACT for the Drive hack: HN 48848840; git-repo pattern is HYPOTHESIS from common practice].
Why current solutions are bad
File-share hacks have no conflict resolution, no shared-vs-private partition (customer notes vs personal scratch), no version history, and no audit trail of which context an agent actually read when it made a decision [FACT: these gaps are enumerated in HN 48848840; audit gap in HN 48846739].
Proposed product
A CLI daemon + small sync server: each teammate's Claude Code memory directory becomes a partitioned store (team/ shared, private/ local-only) synced through a versioned backend (git-like CRDT or last-writer-wins with review queue). Every agent session logs which memory files were injected, producing a replayable audit trail ('what did the agent know when it said X'). Drop-in: works via CLAUDE.md/memory-directory conventions and hooks, no fork of Claude Code required.
MVP version
Self-hostable sync of a designated memory directory with shared/private partition, three-way merge with conflict flags, append-only history, and a session-manifest audit log. Single Postgres/SQLite backend, `curl | sh` install, 2 seats free.
30-day build
Build MVP against founder's own Claude Code setup; recruit the HN 48848840 poster and 5–10 similar teams from that thread and Claude Code communities as design partners; ship self-host alpha.
60-day build
Hosted version with team auth; add the audit/replay viewer (the differentiator vs plain git); publish 'stop syncing your agent's brain over Google Drive' write-up + Show HN.
90-day revenue plan
Convert design partners to $15–25/seat/month hosted plan; target 10–20 paying teams (~$0.5–1.5k MRR) as validation gate before further investment.
Distribution path
Show HN / the two source threads, Claude Code Discord/Reddit communities, GitHub (open-core self-host as top-of-funnel), founder's demonstrated-value style: public demo of replaying an agent's context history. No enterprise procurement needed.
Pricing hypothesis
Open-core: free self-host 2 seats; hosted $15–25/seat/month; audit/compliance tier $50+/seat later. HYPOTHESIS — no pricing evidence in input.
Technical difficulty
Moderate. File sync with partitioning and history is well-trodden (git plumbing/CRDTs); hard edges are merge semantics for prose memory and hooking session-context capture reliably across Claude Code versions. Solo-buildable in 4–6 weeks with AI assistance.
Legal / regulatory risk
Low. Customer data custody (customer notes/strategy) requires baseline security hygiene; self-host option defuses most of it.
Platform dependency
HIGH and the core risk: product lives entirely on Claude Code conventions Anthropic controls and could obsolete with a native team-memory feature. Mitigation: backend-agnostic design (any agent CLI's memory dir), and the audit/replay layer remains valuable even if sync is commoditized.
Founder fit
Mixed. Fits his mold (micro-SaaS, dev-adjacent tool, demonstrated-value selling, fast AI-assisted prototyping) and he is himself a heavy Claude Code operator, so he's his own user. But this is NOT his proven government-portal/forced-buyer shape [lesson, conf 0.80 — does not apply here], the buyer is developers (crowded, taste-driven market), and there's no mandate or deadline forcing purchase. Capital/runway lesson (conf 0.9) applied: not penalised for the ~90-day ramp.
Breakout potential
If agent memory becomes the durable asset teams accumulate, the sync+audit layer is a natural system of record and could expand across agent frameworks; but equally likely the platform vendors absorb it. Moderate.
Final recommendation
VALIDATE-THEN-BUILD, not build-first. The pain is real and well-specified but n≈1 and unpaid; the platform-absorption risk is the highest of any recent candidate. Spend ≤2 weeks: contact the HN poster and 10 similar teams, pre-sell a $20/seat pilot with a mock of the audit-replay view. Build the 4-week MVP only if ≥5 teams commit. Keep it backend-agnostic from day one to hedge Anthropic risk.
Next action
Reply to HN thread 48848840 and DM 10 teams from it and Claude Code communities with a concrete pilot offer + audit-replay mockup; target 5 paid-pilot commitments within 14 days.