localmem Enterprise
The private memory layer for teams and their AI.
localmem reads your documents, meetings and conversations where they already live and turns them into one memory that every AI agent works from.
Built on it today: localmem Teams · SucceedHire · personalchess.club
One memory, any agent
Every agent forgets. Yours don't have to.
Whatever your product runs on, its agents write what they learn to one memory per customer and read back the few facts each answer needs.
- Write what was said, agreed and changed, from calls, chats, notes and documents.
- Understand it once, at write time: who, what, by when, with the words it came from.
- Read the slice that answers the question, as it stands today or as it stood on any date.
- Land it where people work. localmem Teams writes every commitment into your team's Notion.
Claude, OpenAI, Gemini and Cursor are shown as examples of agents that call the API. Their names and marks belong to their owners.
How it works
Capture. Understand. Recall.
- 01
Capture into one log
Every write is added to one event log per customer, encrypted under that customer's keys. Nothing is edited in place.
- 02
Understand into facts
Each capture becomes facts: who, what, by when. Every fact keeps the exact words it came from and a link to its source.
- 03
Keep every version
A changed deadline is a new version. The old one stays, with the date it stopped being true.
- 04
Recall the right slice
Search combines keywords and meaning, then a reranker orders the results. Ask what is true now, or what was known on any past date.
“…fine, I'll send the revised plan by Friday, and let's drop the March launch.”
Maya“Let's give Maya until Monday for the plan.”
Why it is fast and cheap to ask
Other memory hands a model the haystack. localmem hands it the needle.
Most memory feeds the history back into the prompt and makes the model work out again what it already knew, on every call. localmem does the understanding when something is written, so every question gets a precise slice, each fact with its date and its source.
Public evaluations
Measured in the open, on the benchmarks everyone uses.
Memory systems are compared on two public benchmarks. We run all of their questions with the papers' own judge prompts, name the model that reads and the one that judges, and publish every run's records next to its numbers.
- LongMemEval78%*
500 questions across long chat histories: facts, changes over time, and what was true when.
- LoCoMo75%*
Very long conversations between two people, asked about weeks later.
- Search timeWith the full run
From the question to the facts that answer it, on the hosted API.
- Context handed to the modelWith the full run
How much text each answer needs, against feeding the whole history.
* Run on part of each data set. An earlier run on 422 of the 500 LongMemEval questions scored 69%. The full data set run is coming up, and its records will be published next to the numbers.
What it guarantees
Memory you can audit, rebuild and delete.
Each of these is a written requirement of the product, not a line in a slide. The data contract says what is kept, where, for how long and who can read it.
- 01
Keys per customer.
One customer's data is never readable with another's keys.
- 02
History, not overwrites.
A changed deadline or a reversed decision is a new version. The old one is still there, with the date it stopped being true.
- 03
Evidence on every fact.
Each fact points to the quote and the source it came from, so a person can check it in seconds.
- 04
Deletion by destroying keys.
Deleting a meeting destroys its keys. Closing an account deletes its key, which makes every copy unreadable, backups included.
event log · encrypted The log is the record. Search indexes, facts and the graph are rebuilt from it, and a replay reproduces them exactly, without calling a model.
localmem Teams · built on the memory layer
Then it writes it down where your team works.
Every commitment and decision from your meetings becomes a row with its owner, its date, the words that created it and every change since, in your team's own Notion.
Granola
Read AI
NotionWhat it is for
Products that learn from people, and have to stay right about them.
- Customer care
A care agent that knows what each customer was told.
Each customer's setup, history and open promises, as they stand and as they stood.
Read the story - Sales agents
Every deal remembered, call after call.
Each account's people, needs, objections and promises, for AI and human sellers alike.
Read the story - Home care agencies
Every caregiver known, from the first interview.
Skills, availability, preferences and history, for the agents that interview and call caregivers.
Read the story - Medical devices
Every hospital account, remembered.
Each hospital's devices, service history, training and open issues, for field and support teams.
Read the story - Meeting follow-ups
Nothing promised in a meeting gets lost.
A small team's notes become one list of commitments and decisions, kept in Notion.
Read the story - Coaching
A coach that remembers every game.
personalchess.club remembers each player month by month, so its coach knows how a player has changed.
Read the story - Also built on it
SucceedHire: recruiting with voice agents that remember every candidate.
Each screening and follow-up call picks up where the last one left off.
Visit succeedhire.ai
Questions
Common questions
What is localmem?
localmem is the private memory layer for teams and their AI. It reads documents, meetings and conversations where they already live and turns them into one memory that every AI agent works from, with every version kept and each fact pointing back to its source.
How does an AI agent use it?
Your product's backend writes what it learns through the TypeScript or Python SDK, and searches before the model answers, so the model gets the few facts that answer the question instead of the whole history. A coding agent such as Claude Code, Cursor or Codex can set it up from one prompt in the quickstart.
How is it different from a vector database or chat history?
A vector store finds similar text. localmem keeps facts: who, what and by when, each with the words it came from. A changed fact becomes a new version, so you can ask what is true now or what was known on any past date.
Where is the data, and can it be deleted?
Each customer's memory is one event log encrypted under that customer's own keys. Deleting a customer destroys its keys, which makes every copy of its memory unreadable, and backups that still hold it expire within 35 days.
How accurate is it?
On part of each public data set, localmem scores 78% on LongMemEval and 75% on LoCoMo. An earlier run on 422 of the 500 LongMemEval questions scored 69%. The full data set run is coming up.
What does it cost to try?
Test keys are free and reach only test customers. Live plans are on the pricing page, and localmem Teams starts with 14 days free, no card.
Building a product that has to remember people?
Start in test with your own test customers, free. Your coding agent can set it up from one prompt. For a dedicated memory layer, or one in your own cloud, tell us what you are building.