Al Buddy
The Al Buddy mark: a lowercase al

A memory that's yours.

Al Buddy is an open memory for AI assistants. Nothing is dropped to make room, it knows where every fact came from, and it leaves with you.

Al — a name, said like "pal". Not A.I.

What we believe a memory owes you

Most AI memory is a scratchpad the vendor owns: facts get overwritten, nobody can say who claimed what, and you can't take it with you. We built the opposite.

Invalidated, not overwritten

Change is kept, with a date on it

When something stops being true, it gets an end date and stays. You moved cities; your Al still knows where you used to live, and when that changed. Erasing a fact is possible — it is refused unless a policy allows it, and it is written to the audit trail.

Provenance

Every fact says who said it

You, the assistant's guess, a guardian, or the system. And how sure it was. So an assistant can weigh a fact instead of just repeating it.

Portable

It leaves with you

One documented export format with a schema and a proven round-trip. Switch assistants, switch models, switch vendors. Nothing is lost.

Raw text is the truth

Models come and go

Your words are the record. The vectors a model builds from them are a cache: throw them away, rebuild them with a better model, keep everything.

Governed

Rules you can read

Who may write, who may see, what may leave, with an audit trail. Sensitive things stay sensitive. A guardian can protect what needs protecting.

Open

Apache-2.0, one dependency

A single SQLite file on your machine. No account, no key, no service. Read every line on GitHub.

What the others answer

Every agent memory answers "what does it recall?" well. These are the four questions that decide whether the memory is yours — asked of the projects people compare us with. Our reading of their own code and docs, September 2026; if a line is wrong, tell us and we'll fix it.

Question one

Where did this fact come from, and who asserted it?

Here: every fact and every relationship carries it — you said it, the assistant inferred it, a guardian added it, or the system generated it — and it cannot be changed after it is written.

LettaThe git history of its memory files, not a field on a fact.
Mem0A metadata field you can fill in; nothing requires it.
ZepGraph episodes record the conversation a fact came from.
Question two

When was it true, and what replaced it?

Here: a fact that stops being true gets an end date and stays, so "where did he live in 2024?" has an answer. What we do not yet answer: what we believed at a past moment.

LettaFile history, not a model of when a fact was true.
Mem0A change log per memory — old value, new value, when — which is transaction history, not valid time.
ZepIts real strength: graph edges carry valid-from and valid-to alongside created and expired.
Question three

Can I take it with me, losslessly?

Here: one versioned JSON export with a published schema and a round-trip proven by tests. Paste yours into the scorer below and see what survives.

LettaAn agent-state file, shaped like the framework; archival memory is not in it yet.
Mem0A cloud export.
ZepCloud only since 2025; Graphiti, the engine under it, self-hosts.
Question four

Does it run with no vendor, no key, no server?

Here: a SQLite file on your machine and embeddings computed on it. No account, no API key, nothing to sign up for.

LettaSelf-hosting is possible; the cloud is the product.
Mem0Apache-2.0 and local vector stores work, but the default path calls an LLM to extract facts, and the cloud is the product.
ZepZep is cloud; Graphiti self-hosts with a graph database and an LLM key.

Worth naming too, because they answer question four the way we do — on your own machine, no key: basic-memory keeps everything as Markdown notes you own and, alone among these, writes when a note was true (@effective[2026-06-10,2026-07-27)) — what it does not record is who asserted a fact, on a local install; the MCP memory server is the baseline most agents start from, one JSONL file with no source, no dates and a delete that removes the old value; and Memori puts memory in your own SQL database, though extraction needs an LLM key and a fact carries no asserter of its own. Read against their code on 21 September 2026.

Does your assistant's memory know where its facts came from?

Paste any memory export. Seven questions, each scored with its reason. Nothing leaves your browser.

Try a sample:

The rules are public: how each question is scored, which parts are counted off your file and which the adapter declares, and what the score does not measure. Five of the seven are counted from the records you paste, so any system that starts recording provenance scores higher. That's the point.

Governance, in plain words

A memory that plugs into anything needs rules you can read, and an honest line between what the code enforces and what a person still decides.

Enforced in code

What the store will not let happen

A fact is never overwritten — it is closed with an end date and kept. Erasure exists, and a governed store refuses it unless a policy explicitly allows it. Who said a fact, and when, can never be changed after it is written. Sealed facts never surface unless asked for by name. A policy can refuse a write, hide a fact from an audience, or stop it leaving in an export, and every decision lands in an audit trail.

Guardians

Someone can protect what needs protecting

A parent, a carer, a team lead: a guardian's facts can only be changed by a guardian, and everyone can still read them. Built for the people who are looked after, not only the people who look.

Human decisions

Nothing consequential without a person

An assistant on this memory can advise and anticipate freely. Money, deletion, public posts, credentials, new services: those need a human's explicit yes. A false escalation costs less than a false claim of authority.

Honesty

What we do not claim

We do not claim it is safe. We list exactly what is enforced, what is only a prompt, and what is still a person's call, and we change the list in the open. Read the rules.

Use it

For people who build assistants. If that's not you, skip to what's next.

Score an export from the terminal

npx al-buddy-memory conformance my-export.json

Give your agent a memory it can weigh

An MCP server whose every recall returns provenance and validity:

{ "mcpServers": { "memory": {
    "command": "npx", "args": ["al-buddy-memory-mcp"] } } }

Build on the library

npm install al-buddy-memory

TypeScript, SQLite, governance hooks, the portable format. Docs on GitHub.

What's next

The memory is the foundation. What we're building on it:

Soon

Talk to your Al

A real conversation, by voice, with an assistant that remembers what you said last week and can tell you where it learned it. The memory is the part that makes it yours.

Soon

An assistant that learns you

How you decide, what you always ask for, what you'd rather skip. Written down as beliefs with a confidence, correctable in a sentence, never acted on without your yes.

Teams

Memory for teams

The same governance for organisations that need to answer who knew what, and when: audit trail, policies, your own database. Running agents at work and need this? Write to Chris.