Statewave

Statewave's bet is that retrieval is now a commodity and governance is not. The architecture reflects that: raw events land as immutable episodes, a background compilation pass turns them into typed memories with confidence scores, and every context call emits an immutable, ULID-addressable receipt carrying a byte-level integrity hash. That receipt is the piece competitors rarely ship — it lets you replay exactly what an agent saw at a point in time, which is precisely the artifact an auditor asks for. Strengths. The episodic/semantic split is grounded in the 2023 CoALA taxonomy rather than invented ad hoc, and compilation is idempotent, so re-running it never duplicates memory. Sensitivity labels (pii, financial, secret) are auto-detected at ingest but held for operator review before promotion, and the YAML policy engine ships a log_only mode so you can watch access decisions for a while before you enforce them. Multi-tenancy is subject-scoped with app-layer query isolation, and subject erasure clears episodes, memories and receipts in a single call — the GDPR story is a first-class API, not a support ticket. The engineering is documented rather than implied: 708 unit tests and 56 eval assertions across record → compile → retrieve, plus a scoring model the team publishes openly (a resolved session carries a -5.0 ranking penalty; repeat-issue detection adds +4.0, or +6.0 with a summary). Weaknesses. You run it. That means Postgres, deployment sizing, capacity planning, and the tenant-boundary backfill step the team warns silently erases memory if skipped. The connector surface shown in the scrape is GitHub, Slack, Zapier, n8n, Zendesk plus MCP and docs/markdown sources — broad enough for support and coding workflows, but this is not a tool that wires itself into every SaaS you own. The reference benchmark of 73% fewer tokens versus raw history comes from the vendor's own demo, not an independent evaluation. Pricing visible today is a single free Apache 2.0 tier, so budget for your own infrastructure and operations rather than a subscription. Where it fits. Teams running support agents that need session context, handoff packs, health scoring, SLA tracking and repeat-issue detection — the vendor's own "primary wedge" — and coding agents that should accumulate tech stack, architecture decisions and preferences across sessions. Any system where compliance asks for traceability of what was remembered and what was selected into a prompt. Where it doesn't. Non-technical buyers wanting a hosted plug-and-play service, teams that won't operate Postgres, and projects whose recall problem a nearest-neighbor vector lookup already solves. If you need managed convenience more than provenance, Mem0 and Zep are the honest comparisons.