What is Hindsight?
Hindsight is an agent memory server from Vectorize. You send it text with retain, query it with recall, and ask it to reason over what it holds with reflect. Each memory bank is an isolated store for one user, agent or project.
On retain, an LLM (OpenAI gpt-5-mini by default, or any of 25+ providers including Ollama) extracts facts, dates, entities and relationships, which are stored in PostgreSQL with pgvector, or Oracle AI Database 23ai, as entities, links, time series and dense and sparse vectors. Recall runs semantic, BM25 keyword, graph and time-range searches in parallel, merges them with reciprocal rank fusion and reranks with a cross-encoder. In the background, related facts are consolidated into observations that keep their supporting quotes and are refined rather than overwritten when new evidence arrives.
On 3 November 2026 Vectorize released three small open-weight models for the memory pipeline: a 0.8B decision model, a 114M reranker and a 112M embedder. The server is MIT-licensed; Hindsight Cloud is the paid hosted version.
Is Hindsight worth it?
Hindsight has grown faster than any rival: about 59,400 GitHub stars, weekly releases and v1.2.0 on 3 November 2026. Unlike Mem0, it keeps its full feature set in the MIT-licensed server, including graph and temporal retrieval, background consolidation and an MCP endpoint per bank.
The design treats memory as beliefs, not a pile of chunks. Retain uses an LLM to pull out facts, entities, relationships and dates; recall runs four searches at once (vector, BM25, entity graph and time range) and fits the result to a token budget that defaults to 4,096. Observations are the distinctive part: related facts are merged into a belief that records its evidence and a proof count, and new information strengthens or weakens it instead of silently replacing it. Mental models go a step further, storing a standing answer to a question such as a user's preferences, which is read back with no retrieval and no LLM call.
The benchmark claims are strong but not independent. Vectorize's paper reports 91.4% on LongMemEval and 89.61% on LoCoMo with Gemini-3 writing answers, and 83.6% on LongMemEval using only the open GPT-OSS-20B model. Its live leaderboard now shows 94.6% on LongMemEval and 92% on LoCoMo. The README says the results were reproduced by Virginia Tech and The Washington Post, but researchers from both are co-authors of the paper, and the competitor figures in it were copied from other vendors' reports rather than rerun. No retrieval latency is published.
The self-hosted server is not locked down by default. API key checks only start once you set HINDSIGHT_API_TENANT_EXTENSION, and every bank gets an MCP endpoint, enabled by default, so a server exposed on port 8888 lets any caller read and write memory. Secret and PII scrubbing (Memory Defense, 45 patterns) is opt-in per bank. Vectorize's pricing page says the self-hosted build sends no telemetry.
The hosted bill is driven by writes. Retain costs $10 per million tokens, five times Honcho's ingestion rate, while recall is $0.75 per million; storage is free for 30 days and then $0.25 per million tokens a month, and the pricing page says memories are never deleted. An agent that stores 100 million tokens of conversation a month pays $1,000 before any recall.
Where it earns its place
A platform engineer giving a 30-person team's coding agents shared project memory
Stop Claude Code and Codex relearning the same repo conventions every session
npx @vectorize-io/hindsight-coding-agents install all wires each detected agent to a per-repo bank built from git history and past sessions. Self-hosted on one Postgres box it costs only the extraction LLM; put it behind the API key extension before sharing it on the network.
A product team running a support assistant for a 50,000-customer subscription business
Remember each customer's history and notice when a fact changes
One bank per customer, with retain after each ticket and recall before each reply. Observations keep the old and new plan with their evidence rather than returning both blindly. On Cloud, 100 million retained tokens a month is $1,000; self-hosting moves that to LLM extraction costs.
An ML engineer at a bank that cannot send customer text to a hosted API
Run agent memory entirely on its own hardware
Point Hindsight at a local model through Ollama or vLLM and run the open hs-decide-1-flash, hs-rerank-1 and hs-embed-1 models locally to cut LLM calls. The paper's 83.6% LongMemEval result used only GPT-OSS-20B, which fits on one high-end consumer GPU.
Before you commit
- Setup
- One docker run with an LLM key starts the API on port 8888 and a UI on 9999 with an embedded Postgres; pip install hindsight-all runs it in-process from Python. Helm charts and an external-Postgres Compose file cover production.
- Integrations
- Python, TypeScript and Go clients, a CLI and REST API; an MCP endpoint per bank; an LLM wrapper for OpenAI and Anthropic SDKs via LiteLLM; 60+ integrations including LangGraph, CrewAI, Pydantic AI, n8n and Claude Code.
- Data and privacy
- Self-hosted data stays in your Postgres or Oracle database, but retained text goes to the configured LLM; Vectorize says the self-hosted build has no telemetry. Cloud keeps daily backups, says memories are never deleted, and shows SOC 2 Type 2 badges; BYOC, single-tenant, SSO and Memory Defense are Enterprise features.
- Cost at scale
- Self-hosting costs Postgres plus extraction and consolidation LLM calls, which the open decision model is meant to reduce. Cloud: $10 per million tokens retained, $0.75 per million recalled, $0.05 per reflect, $0.25 per million tokens stored a month after 30 days.
- Limits
- Retrieval latency is not published; consolidation is asynchronous, so new facts reach observations with a delay. 185 open issues and pull requests, and a fast release cadence means frequent upgrades.
What’s good
- Full feature set under MIT, including graph and temporal recall
- Observations revise beliefs with evidence instead of overwriting
- Runs on Postgres with an embedded option, or fully local with Ollama
- Built-in MCP endpoint and plug-ins for 13 coding-agent CLIs
What’s not
- API authentication off until you configure a tenant extension
- Benchmark reproduction by the paper's own co-authors
- Cloud retain at $10 per million tokens is the priciest write here
- Every retain and consolidation pass costs LLM calls
Best for
Teams that want graph, temporal and vector memory in one self-hostable server without an open-core feature split.
Hindsight alternatives
Hindsight: common questions
- Is Hindsight worth it?
- Hindsight is the most capable open-source agent memory server you can run today, with nothing held back from the MIT licence, though its benchmark lead rests on numbers its own paper's co-authors reproduced. We rate it 4.4 out of 5.
- How much does Hindsight cost?
- Self-hosted server free under MIT. Hindsight Cloud is usage-billed with $5 of free credit: retain $10 per million tokens, recall $0.75 per million tokens, reflect $0.05 per call, storage $0.25 per million tokens a month after the first 30 days. Enterprise (BYOC, on-premises, SSO) by quote.
- What are the downsides of Hindsight?
- API authentication off until you configure a tenant extension; benchmark reproduction by the paper's own co-authors; cloud retain at $10 per million tokens is the priciest write here; every retain and consolidation pass costs LLM calls.
- Who is Hindsight best for?
- Teams that want graph, temporal and vector memory in one self-hostable server without an open-core feature split.
- How much does Hindsight cost for a team or in production?
- Self-hosting costs Postgres plus extraction and consolidation LLM calls, which the open decision model is meant to reduce. Cloud: $10 per million tokens retained, $0.75 per million recalled, $0.05 per reflect, $0.25 per million tokens stored a month after 30 days.
- Is Hindsight safe for company data?
- Self-hosted data stays in your Postgres or Oracle database, but retained text goes to the configured LLM; Vectorize says the self-hosted build has no telemetry. Cloud keeps daily backups, says memories are never deleted, and shows SOC 2 Type 2 badges; BYOC, single-tenant, SSO and Memory Defense are Enterprise features.
- What are the best Hindsight alternatives?
- The closest alternatives we have reviewed are Mem0, Graphiti, Honcho, Letta and Memvid.
Answers checked Nov 6, 2026.
Make Hindsight? Show this rating on your site
