Note: unprocessed — watchlist.
Why this is interesting
Hyperspell is pitching a “synthesised company brain” — they continuously ingest Slack, email, Drive, CRM, calendar, and meeting transcripts, do conflict resolution + identity unification + source-authority ranking + freshness tracking, and surface the resulting context graph as a filesystem any agent can read. Same category as Glean, Onyx, Coral, Zep’s team graph, but leaning heavier on synthesis than search.
Two points their pitch makes that land independently of the product:
- Retrieval at runtime is a scavenger hunt. Starting from zero every query is wrong; you want a pre-synthesised layer. This exactly mirrors the direction in our shared project memory research — committed knowledge compounds, RAG-on-demand does not.
- Filesystem-as-delivery is the right interface. Every agent already reads files. That’s what our
status/sessions/,research/,decisions/folders already do for a single-team single-repo case.
Why we didn’t adopt it now
Current view from the Multica + Managed Agents research: Hyperspell is solving the 50–500-person-company version of a problem 361 has at 5-person scale. Source fragmentation at 3–5 devs on one product is low; identity unification is trivial (TEAM.md does it in 20 lines); privacy exposure of handing Slack / email / Drive to a managed service is a material NZ Privacy Act concern for tender-sensitive data; pricing undisclosed; their own benchmark admits “early results are humbling, for everyone.”
Proposed research angle (for future processing)
- Is there an OSS or self-hostable Hyperspell variant? What’s the commercial licence?
- Actual pricing for a 5–10-seat team.
- NZ / ap-southeast-2 data residency story.
- Independent evaluations of the context graph (any Context Bench results yet).
- How it compares to
garrytan/gbrainand the LLM-wiki derivatives already surveyed in the memory note. - Revisit triggers: team grows past ~8 people; source fragmentation across multiple active client projects; Hyperspell ships a self-hostable tier.
Related in-repo docs
- Shared project memory research — the memory-layer decisions this would plug into.
- Multica + Claude Managed Agents research — the orchestration layer that would feed it.
- Database analysis — data-tier context.