Give your agents a past
Centralize all your information in a single place.
Expose that information to AI agents in the best possible way.
Your information is scattered across your life or your organisation: email, chat, documents, meeting notes, tickets, agent transcripts. 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.
RememberStack is memory for that agent. You give it the documents your work already produces. 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 and what was true a month ago.
RememberStack is the open-source memory engine. remember is its Python
client and CLI.
New to memory systems? Start with What is a memory system?, written for any reader.
What RememberStack keeps that a context window loses
What each source said. Every statement worth keeping is stored as a claim, together with the exact passage it came from and the date the source said it. Claims are never edited. If a spec from March said the billing migration would ship in June, that stays on record even after the plan changes.
What is true now. From those claims, RememberStack maintains facts about the people, systems and decisions in your work. When a newer source changes a fact, the fact changes, and the claim it replaced 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. You can ask what is true today, what was true on a given date, what held during a quarter, or how something changed over time.
Where it came from. Every fact links to the claims that support it, and every claim to the characters in the source document. Your agent can quote and cite. You can check.
What your agent gets back
When your agent asks a question, RememberStack does not return the passages most similar to the question. It returns 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.
If the memory has nothing on the subject, the result says so. It does not hand back the closest text it could find. If a name could mean two different people, it says that too. An agent that is told "unknown" stops inventing.
No language model writes the answer. The only model involved in reading is the embedding model that turns your question into a search vector. The same question against the same memory returns the same result, and you can inspect exactly why each item is there.
Why not a vector database
A vector search finds the passages that sound most like your question. It cannot tell you which of them is still current, which one a later document corrected, or whether two of them contradict each other. It has no idea when anything was true. And it always returns something, even when nothing relevant exists.
| 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). Claims are the transcript; facts are the verdict. |
| A correction overwrites what came before | The old fact's time window is closed, not erased. You can still ask what held before. |
| Contradictions are averaged away or hidden | Both sides come back, marked as a contradiction. |
| Re-sending a file makes it look more certain | Support counts distinct documents. Re-processing a file, editing it or repeating a sentence does not add support. |
| The search index is the authority | Indexes only nominate candidates. The database confirms each one against what is currently held before it is returned, and the result says how many were dropped. |
| A model writes the answer at query time | No language model writes the answer. Your agent plans; RememberStack returns what is known, with its evidence. |
| "No results" could mean anything | The result says which: the entity is unknown, or it is known and nothing matches. |
RememberStack still uses vector search, as one signal among several. The others are keyword search, the graph of how entities relate, and time.
Where it runs
RememberStack is open source under Apache-2.0 and runs on your own infrastructure. Everything it does with your documents is in the public repository. Requirements says what it needs.
Start here
- What is a memory system?: the idea in plain language, for any reader.
- A five-minute tour: what happens to a document from the moment you send it to the moment your agent asks about it.
- Quickstart: send a document and ask your first question.
- Connect your coding agent: give Claude Code, Cursor, Codex or Claude Desktop access to your memory.
- Time: the idea that sets RememberStack apart from retrieval over text.