← All comparisons
ValorBrain · vs · Hindsight
ValorBrain vs Hindsight
Hindsight is the memory product from Vectorize.io, creators of the Agent Memory Benchmark (AMB). Biomimetic architecture with three memory types (World, Experiences, Mental Models) and four parallel retrieval strategies: semantic, keyword, graph, and temporal. The retrieval core is closed in a Docker image; the MIT repo only opens clients and CLI.
agent memory with temporal reasoning (Vectorize) · vectorize.io
Feature matrix
Focused on multi-tenant team memory. Not generic chatbots.
| Capability | Hindsight | ValorBrain |
|---|---|---|
| Multi-tenant Postgres RLS | — | yes |
| Per-tenant Filtered DiskANN | — | yes |
| Reader-agnostic (zero LLM in retrieval) | — | yes |
| Open-weights models | — | yes |
| Temporal search leg | yes | on roadmap |
| Reflect / automatic consolidation | yes | consolidation active, reflect on roadmap |
| Correctability (keyed_facts + authority) | — | yes |
| Queryable knowledge graph | internal, closed | yes |
| Public MCP HTTP transport | — | yes |
| OAuth 2.1 dynamic registration | — | yes |
| On-prem self-hosted | Docker with closed image | yes |
| BEAM-100K · RAG average score (different configurations) | 86.2% (reader Gemini 3.1 Pro, judge Gemini 3.5 Flash) | 75.5% (reader GLM-5.3 Flash (Ox Alpha); judge GLM-5.2; effort=max; run 2026-08-23; ~44% slower than default effort; this benchmark setup is not the chat default) |
| Independent third-party validation | yes | — |
| Per-person permission on every document | — | yes |
| Priced per person, not per call | per query | yes |
Where Hindsight shines
- Four retrieval legs (semantic + BM25 + graph + temporal) — the temporal leg is unique in the market
- Time series as first-class structure: attribute value over time is queryable
- "Reflect" operation: analyzes memories to form connections and synthesize higher-order knowledge
- Independent validation: 73.4% on BEAM single-query, reproduced by Virginia Tech and The Washington Post
- Massive open-source community (19K+ stars) and academic paper on arXiv
- Mature Python/Node SDKs and 8 LLM providers supported in the product
ValorBrain edge
- Native multi-tenant Postgres RLS with per-tenant Filtered DiskANN
- Reader-agnostic retrieval: zero LLM calls to search and rank; reader, effort, and judge are disclosed separately for every BEAM run
- Open-weights models across the pipeline (LFM2.5, BGE-Reranker, GLiNER) — no vendor dependency
- Correctability: keyed_facts with as_of, authority levels (human > designated > agent), temporal supersede
- Open Postgres extensions: pgturbohybrid, pg_ripple, pg-trickle, pg-deltax
- Real on-prem: runs on the client network, zero calls leave
- Single /api/v1/memory/prepare endpoint — same endpoint for production and benchmark
Gaps for multi-tenant SaaS
- Single-tenant: no native Postgres RLS multi-tenant isolation
- Closed core: the MIT repo does not contain the retrieval service (proprietary Docker image)
- Structural Gemini dependency: the AMB benchmark they created is hardcoded to Gemini as default
- No correctability: memory is write-once, no authority levels or temporal supersede
- No per-person permission on each document
- No public MCP HTTP transport with OAuth 2.1
When to choose each
Choose Hindsight if…
- Temporal reasoning and event ordering is your absolute priority
- You want academic validation and a large open-source community
- You already use Gemini and want home-court advantage in the benchmark
- Single-tenant is sufficient (one agent, one memory)
Choose ValorBrain if…
- Multi-tenant B2B SaaS with isolation between companies
- You want reader-agnostic memory that works with any LLM
- On-prem is a requirement (data cannot leave the network)
- You need correctability: facts that change, human authority, temporal supersede
- Multiple agents need the same coherent memory
Try the memory layer
Free credit on signup. Connect a source and ask with citations.