Two agent memories.
Different centres of gravity.
Hindsight, from Vectorize, is an MIT-licensed memory engine for agents (retain, recall, reflect). Brain is shared memory for teams of agents and people. Both run the same core retrieval recipe. Where they differ is below. Hindsight details come from its site and paper, reviewed between Aug and Oct 2026.
Where Hindsight is ahead
| Capability | Hindsight | Brain |
|---|---|---|
| Time-aware search | Parses dates out of the query ("last spring" becomes a window). Stores when a fact happened separately from when it was mentioned. Time is its own retrieval channel. | Ranks fresher documents higher and lists documents by date window. Search itself takes no time filter, and queries aren't date-parsed. |
| Entity graph in retrieval | Spreading activation over the entity, causal and temporal links pulls in related facts the query didn't name. | Has an entity graph you can browse and query. It doesn't feed search ranking. |
| Belief confidence | Opinions carry a 0–1 confidence that new evidence raises or lowers. | Facts are active, superseded, disputed or retired. No per-belief score. |
| Deployment and reach | Cloud, embedded, or your own Postgres. 51.4k GitHub stars (per its site, Oct 2026). | Hosted, or on-premise. |
Both do
| Capability | How |
|---|---|
| Hybrid retrieval | Keyword and semantic results fused (RRF), cross-encoder rerank, results cut to a token budget. |
| Consolidated summaries | Per-entity rollups that keep history: the newer fact stated as current, the older one as "previously". |
| Supersede and validity | A replaced fact is marked invalid from a point in time, not silently overwritten. |
| Near-duplicate merging | Near-identical facts are detected and merged or kept by an LLM check. |
| Isolated stores | Hindsight calls them banks; Brain calls them vaults. |
Only in Brain
| Capability | Brain | Hindsight |
|---|---|---|
| Team memory | Grant other users access to a vault. | Banks are single-owner. |
| Agent-to-agent handoffs | Agents send each other tasks through a durable inbox. | None. |
| Lineage and trust | Every fact traces to its source documents, with a trust tier and origin. Runbooks are marked draft or ratified. | Facts carry evidence and provenance; no trust tiers or ratification step. |
| Obsidian | Two-way sync with a vault. | One-way markdown mount of its knowledge pages. |
| Secret redaction | Always on at ingest, for every vault. | Opt-in per bank. Covers PII too and can block, not only redact. |
| Work continuity | Sessions, todos and skills live next to the memory. | Memory and knowledge pages; no todos or skills. |
Where Hindsight fits
One agent or one user that needs strong recall across time: questions like "what changed since March", or memory embedded inside your own app. It's MIT-licensed and self-hosts in one container.
Where Brain fits
Several agents and people working from the same memory: shared vaults with access grants, handoffs between agents, and facts you can trace back to their source before you act on them.
Benchmarks. Hindsight reports LongMemEval-S 94.6 and LoCoMo 92.0 on its site. Those are its own numbers, scored by judge models that differ from the baselines it compares against. Brain hasn't published scores under a third-party harness, so we don't put the two side by side.
Method. Hindsight: repo vectorize-io/hindsight and paper arXiv:2512.12818 (reviewed 19 Aug 2026), site re-read 25 Sep and 19 Oct 2026. Brain: checked against its source on 19 Oct 2026. Both projects move fast; check the current docs.