What Hindsight actually stores
Hindsight is a memory server you put next to an agent. Its site puts the design choice plainly: “Hindsight does not store conversations.” When you send it a conversation or a document, a language model pulls out individual facts, gives them a time range where possible, links them to the people, companies and things they mention, and stores them in Postgres with pgvector.
The research paper behind it, “Hindsight is 20/20” (December 2025, by authors from Vectorize.io, Virginia Tech and The Washington Post), sorts those facts into four networks:
| Network | What it holds | Example |
|---|---|---|
| World | Objective facts about the outside world | “The client’s fiscal year ends in March” |
| Experience | What the agent itself did and experienced | “I sent the draft on Tuesday and got edits back” |
| Opinion | Subjective judgments with a confidence score that can change | “This client prefers short summaries (confidence 0.7)” |
| Observation | Neutral summaries of an entity, built from the facts | A profile of one customer |
The current product site uses slightly different words: world facts, experience facts, observations, mental models and pages. A mental model is a standing answer to a question such as “what are this user’s preferences?”, which Hindsight keeps up to date in the background so reading it needs no LLM call.
Retain, recall, reflect
Everything goes through three calls, available from Python, TypeScript, Go, a CLI, HTTP and MCP:
- Retain stores what happened. Extraction, entity linking and graph building run behind this call.
- Recall searches it back. Four searches run in parallel: semantic (vectors), BM25 keywords, graph (following entity links) and temporal (matching time ranges). The results are fused by rank, reranked by a cross-encoder and cut to a token budget, not a fixed number of results.
- Reflect reasons over the bank. It recalls memories and then writes an answer shaped by the bank’s background and disposition, and it can form new opinions along the way.
Memories live in banks. The README describes a bank as an isolated store, meant as one “brain” for one user, agent or project, with no leakage between banks. Each bank gets its own MCP endpoint at /mcp/{bank_id}/, switched on by default.
How to run it
The quick start is a single Docker container that exposes the API on port 8888 and a UI on port 9999, with your LLM key passed as HINDSIGHT_API_LLM_API_KEY. For production there are Docker Compose with an external Postgres, Helm for Kubernetes and a pip install. More than 25 LLM providers can do the extraction, from OpenAI and Anthropic to local models through Ollama or llama.cpp. If you do not want to host it, Hindsight Cloud is the managed version. The code is MIT-licensed.
The misunderstanding about the benchmarks
Hindsight’s site shows high scores, for example 94.6% on LongMemEval-S and 92.0% on LoCoMo, against an unnamed “next best” system. Those are vendor-reported numbers. The paper reports lower figures for its own setups (83.6% on LongMemEval with a 20B open model, 91.4% with Gemini 3 Pro writing the answers) and notes that the authors could not reproduce a competitor’s claimed LoCoMo score. The paper also says the benchmarks test factual recall, not the opinion and disposition features that make Hindsight different. So the scores tell you Hindsight retrieves facts well in those tests; they do not tell you it will remember your team’s way of working. How these benchmarks are built is explained in How do you evaluate LLM agent memory?
What this means for a team
Hindsight is built around banks per user or per agent, which is the right default for privacy. For a team it means you decide the layout yourself: one bank per person keeps knowledge apart, one shared bank mixes everyone’s facts and opinions together, and the README mentions tenant and auth extension points but gives little detail. You also run and secure the server, and every write passes through an LLM that can store something wrong, which is a reason to read What is memory poisoning in AI agents? before you let agents write freely. For a comparison with Mem0, Zep and Letta, see Mem0 vs Zep vs Supermemory vs Letta.
Where Bombos fits
Hindsight is a strong building block if you want to wire memory into your own agent and run it yourself. Bombos starts from the team instead: one shared memory for every AI tool the team already uses, from Claude and ChatGPT to Copilot, Cursor and your own agents, with team knowledge, learned rules about how you work and the knowledge already in mail, documents and repos, and you can see what each answer rests on. It runs in our cloud or on your own server, and one person can try the plugin first for their own agents.
Related questions
Is Hindsight free to use?
The code is open source under the MIT license, so self-hosting costs only your own infrastructure and the LLM calls it makes. Hindsight Cloud is a managed version; its site mentions free credit at sign-up but lists no price tiers.
Does Hindsight need an LLM?
Yes. Extracting facts, resolving entities and running reflect all use a language model. The Docker quick start asks for an LLM API key, and the README lists more than 25 providers, including local options such as Ollama and llama.cpp.
Can I use Hindsight with Claude Code or Cursor?
Hindsight's site lists integrations for more than 20 coding harnesses, including Claude Code, Codex CLI and Cursor CLI, and every memory bank exposes an MCP endpoint with retain, recall and reflect as tools.
Are Hindsight's benchmark scores independent?
The scores on its site and in its paper are reported by the authors, most of whom work at Vectorize. The paper itself notes it could not reproduce one competitor's LoCoMo claim, which shows how hard these numbers are to compare.
