RememberStack

Give your agents a past.

RememberStack is open-source memory for AI agents. You give it the documents your work already produces: email, chat, specs, meeting notes, tickets, agent transcripts. It reads them, keeps what matters, and answers your agent's questions with facts it can trace back to the source. It knows what is true now, what was true a month ago, and who said so.

Read the docs · Quickstart · What is a memory system?

---

Why you should care

Every AI session starts from zero. The agent does not know which decision still holds, which file said it, or what changed overnight. You paste the same notes in again, or the model fills the gap with something plausible.

Most "memory" for agents is a vector search over text: it finds passages that sound like the question. It cannot tell you which of them is still current, which one a later document corrected, whether two of them contradict each other, or when anything was true. And it always returns something, even when nothing relevant exists.

RememberStack keeps four things a context window loses:

• What each source said. Every statement worth keeping is stored as a claim, with the exact passage it came from and the date it was said. Claims are never edited.
• What is true now. From those claims it maintains facts about the people, systems and decisions in your work. When a newer source changes a fact, the fact changes and the old one stays on record. When two sources disagree, you see both, marked as a contradiction.
• When it was true. Facts carry the period in which they held. Ask what is true today, what held last March, or how something changed.
• Where it came from. Every fact links to its claims, every claim to the characters in the source. Your agent can quote and cite; you can check.

What your agent gets back

Not the passages most similar to the question, but the facts that answer it, each with its time window, the evidence behind it and any contradiction, in one typed result that is ready to put into a prompt.

• Honest negatives. If memory has nothing on the subject, the result says so. If a name could mean two different people, it says that too. An agent that is told "unknown" stops inventing.
• No model writes the answer. The only model on the read path embeds your question for search. The same question against the same memory returns the same result, and you can see why each item is there.
• Search nominates, the database confirms. Vector, keyword, graph and time signals propose candidates; every one is checked against what is currently held before it is returned.

Where typical agent memory goes wrong	What RememberStack does instead
Source text is treated as the truth	What a source said (a claim) is kept apart from what is held true now (a fact)
A correction overwrites what came before	The old fact's time window is closed, not erased
Contradictions are averaged away	Both sides come back, marked as a contradiction
Re-sending a file makes it look more certain	Support counts distinct documents, not repetitions
"No results" could mean anything	The result says whether the entity is unknown or known with nothing matching

Built for agents

• MCP. remember mcp gives Claude Code, Cursor, Codex, Claude Desktop and other MCP clients the memory tools; remember setup configures them in one command. Connect your agent →
• Python SDK and CLI. pip install remember. SDK → · CLI →
• HTTP API. The same operations over plain HTTP. API →
• Your files. Markdown, text, HTML, Word, PowerPoint and Excel convert out of the box; PDFs and images with an OCR converter. File types →

Quickstart

You need Docker (Engine 28 or later for the tokenless local setup) and an OpenRouter API key.

git clone https://github.com/writeitai/remember-stack && cd remember-stack
cp .env.example .env
printf 'REMEMBERSTACK_POSTGRES_PASSWORD=%s\nREMEMBERSTACK_MINIO_ACCESS_KEY=%s\nREMEMBERSTACK_MINIO_SECRET_KEY=%s\nREMEMBERSTACK_SELFHOST_DEPLOYMENT_ID=%s\n' \
  "$(openssl rand -hex 32)" "$(openssl rand -hex 12)" "$(openssl rand -hex 32)" \
  "$(openssl rand -hex 16 | sed -E 's/^(.{8})(.{4}).(.{3}).(.{3})(.{12})$/\1-\2-4\3-8\4-\5/')" >> .env
# edit .env: set REMEMBERSTACK_OPENROUTER_API_KEY
docker compose up -d

uvx remember setup --self-hosted   # connect your coding agent

Then send a document and ask your first question: Quickstart → · Self-hosting →

Open source, all of it

Apache-2.0. Everything that decides what your memory holds (extraction, entity resolution, supersession, provenance, forgetting) is in this repository. Nothing that affects correctness is held back.

• Docs: remember.dev/docs
• Current release: v0.17.2 · PyPI · container
• How it works inside: architecture · design corpus · decision log

Contributing

See CONTRIBUTING.md and CLA.md. Pull requests need the contributor-agreement checkbox in the PR template.