Hindsight by Vectorize gives agents memory that survives between sessions: per-user context persisting for weeks, recall in under 100 milliseconds, and a reflection layer that synthesises observations into consolidated knowledge rather than accumulating raw facts. It works with any LLM and exposes remember, recall and reflect through MCP.
The reflection layer is the design decision that matters. A memory store that only appends becomes a pile nobody can retrieve usefully from, which is why most agent memory degrades into a slow and noisy vector search; consolidating observations into knowledge is what a person does between sessions and what makes recall useful at scale. Learning from tool failures and agent mistakes is the same idea applied to errors.
Nvidia, Groq and Electronic Arts are named as users. Hindsight Cloud exists but no pricing figures are published. Persistent per-user memory is personal data with retention and deletion obligations attached, a reflection layer makes editorial choices about what matters, and memory that survives weeks also preserves things that have since changed.
