Skip to main content

RememberStack

CI Docs PyPI License

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.

Claims (what sources said) feed facts (what is true now)

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

Search returns similar passages; memory returns what is known

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 --build

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.

Contributing

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

Metadata

Release files for remember 0.19.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for remember 0.19.1
File Size Uploaded Provenance Actions
remember-0.19.1.tar.gz 7.9 MB With provenance Details for remember-0.19.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for remember 0.19.1
File Interpreter ABI Platform Actions
remember-0.19.1-py3-none-any.whl Python 3 none any Details for remember-0.19.1-py3-none-any.whl

Total release size: 8.1 MB

Release files / remember-0.19.1.tar.gz

Download URL remember-0.19.1.tar.gz
Size 7.9 MB
Tags Source
SHA-256 checksum
How to use checksums
bcb0416dd358cee565620301a0550efbc6f075af10a5734cd099111d6a51eca9
BLAKE2b-256 checksum
How to use checksums
70cd049f2c6d2345d95ed2eb103af109a17a9776c629aa86352b1fc889276b83
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 11, 2026.

Transparency log

Release files / remember-0.19.1-py3-none-any.whl

Download URL remember-0.19.1-py3-none-any.whl
Size 138.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
5a25c329dadd6ee049f83b071de777f74a8e61692db04c908d9f6f8354ae5661
BLAKE2b-256 checksum
How to use checksums
2d174c3d849912449b823a5e265eb64f1a261978913e7be10ca1f4670d7f5639
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 11, 2026.

Transparency log

Security for remember 0.19.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for remember-0.19.1.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 11, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page