35,388 stars · MIT, read from /blob/main/LICENSE on 2026-09-28; standard unmodified text, copyright "2025 Vectorize AI, Inc.". THE PROJECT'S OWN CHANNELS DISAGREE: the repository and the npm package say MIT, the PyPI page for hindsight-api says Apache 2.0 · v0.10.1 (2026-09-21), read from /releases/latest and confirmed by both ungh and PyPI; only two releases exist in total · Track this in Scout
A memory store for AI agents that keeps what mattered from past sessions and hands the useful parts back, rather than replaying a whole conversation history.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A server plus client libraries for Python and JavaScript that give an agent long-term memory. It stores memories in PostgreSQL with the pgvector extension and uses a language model to decide what is worth keeping and what to hand back.
What it is good for. Anyone building an assistant that people use more than once. A support assistant that remembers a customer's setup, a coding assistant that remembers the house rules, a personal agent that stops asking the same three questions. It is developer infrastructure, so it is for people writing the application rather than using one.
- MIT, read from the LICENSE file, with no restrictions at all — unusual in a category where most projects are commercially backed.
- It works with more than twenty-five model providers, so nothing is locked to one vendor.
- PostgreSQL rather than a special database, which means backups and monitoring you may already know how to do.
- It has an unavoidable running bill. Every memory written and every recall calls a language model and an embedding model, so the cost grows with how much the agents remember and re-read. No pricing guidance is published for people running it themselves.
- Only two releases exist — v0.10.0 on 14 September 2026 and v0.10.1 on 21 September 2026 — and the newest tag is already 140 commits behind the main branch. Pinning a release means running well behind the code the maintainers are testing. The JavaScript client is at 0.8.4, two minor versions behind the server.
- Its own publishing channels disagree about the licence: the repository and the npm package say MIT while the PyPI page says Apache 2.0. Both are permissive, so the practical risk is small, but it is a contradiction in the project's own metadata.
mem0ai/mem065,967 stars, code on 25 September 2026, covered in Edition 28; the leader in this category, with a paid hosted service alongside it, and a more general store rather than one built around learning from experience.
Track this in Scout- getzep/graphiti
31,139 stars, code on 24 September 2026; stores memory as a knowledge graph in a graph database rather than in PostgreSQL, which suits questions about how things relate.
Track this in Scout - letta-ai/letta
24,870 stars, code on 10 September 2026; a whole platform for building agents inside, rather than a memory component added to an agent that already exists.
Track this in Scout
# Python pip install hindsight-all -U # JavaScript npm install @vectorize-io/hindsight-client
