Memory that infers,
updates and forgets.

You hand Esfis what your assistant sees: conversations, files, URLs, notes. It turns them into facts, connects the facts to each other and to time, and gives them back as a profile, a search result, or a file the agent can grep.

What a memory is

Documents are what you send. Memories are what Esfis keeps.

	Documents	Memories
What	The raw input you send	The facts Esfis extracts
Examples	A chat log, a PDF, a meeting note, a URL	“Alex is a PM at Stripe”
Role	Source of truth for retrieval	Personal and entity state over time
Lifecycle	You add, update and delete	The graph updates, extends, supersedes and forgets

One input becomes many memories. Each is atomic, carries enough context to stand alone, and is linked to what the graph already knew about that person, that project, that place. You define no schema and draw no edges.

ONE MESSAGE IN
Had a great call with Alex. He's enjoying the new PM role at Stripe,
though the payments work is intense. He moved to Seattle for the job,
Capitol Hill. Wants dinner next time I'm in town.

FOUR MEMORIES OUT
Alex works at Stripe as a PM                    · since 2026-08
Alex finds the payments work intense            · 2026-09-04
Alex lives in Capitol Hill, Seattle             · supersedes: Alex lives in Austin
Alex wants dinner when I'm next in Seattle      · open intent

Infers, updates, forgets

Infer

Facts are extracted with the surrounding context and a timestamp. Implicit information counts: “moved for the job” becomes a location fact and a career fact.

Update

When a new fact contradicts an old one, the old one is superseded, not overwritten. The history stays, so you can ask what you knew as of any date.

Forget

Facts decay by relevance and can be pinned or deleted by hand. Deletion is a commit too: audited, reversible until you purge.

Memory blocks

Some things should never depend on retrieval. Blocks are named, evergreen documents injected ahead of recall on every request: who you are, how you like to be answered, what you are working on this quarter.

~/.esfis/blocks/
├── persona.md      # how you want assistants to talk to you
├── facts.md        # stable facts: name, city, timezone, employer
├── projects.md     # what you're working on, updated by the observer
└── team.md         # who's who, shared with your Team space

Edit them from the app, from any editor, or let your assistant maintain them for you. Blocks are versioned like everything else.

Recall

Recall runs in four deterministic stages. Every stage is inspectable, and the whole cascade completes in under 50 ms on a laptop.

Keyword
Full-text search over facts and documents. Exact names, dates and identifiers win here.
Meaning
Local embeddings over the same set. Catches paraphrase and intent.
Graph
Expand one hop through linked entities and time. “Dinner in Seattle” pulls in Alex, and Alex pulls in Stripe.
Rerank
A small cross-encoder orders the candidates and trims to a token budget you set.

GET /recall?q=what+should+I+cook+for+Friday&budget=800

200 OK
[
  { "fact": "Dinner with Laura & Dani, Friday 9pm", "score": 0.94, "as_of": "2026-09-08" },
  { "fact": "Laura went gluten-free this spring", "score": 0.81, "as_of": "2026-04-12" },
  { "fact": "Dani doesn't eat shellfish",          "score": 0.77, "as_of": "2025-11-30" }
]

Try it

A sample store of twelve memories, the real cascade, running in your browser. Ask it something.

Keyword3 ms

Meaning12 ms

Graph8 ms

Rerank20 ms

What feeds it

Assistants

Claude, ChatGPT, Cursor, Perplexity and any MCP client, through the observer.

Files

Markdown, PDF, plain text, email exports. Drop a folder, it stays watched.

URLs

Articles and docs you save. Readable text only, images ignored.

API

POST anything your own agents see. Same pipeline, same store.

Where it runs

Surface	What you get	Status
macOS app	Menubar engine, memory browser, block editor, audit timeline	Stable
MCP server	Local stdio or remote HTTP with OAuth 2.1	Stable
CLI	`esfis recall`, `esfis log`, `esfis diff`	Stable
Browser plugins	ChatGPT and Perplexity injection	Beta
Linux / Windows	Engine and CLI, no app yet	Beta
iOS	Read and pin memories on the go	Coming

Our products

Memory Engine

The core: temporal knowledge graph, cascade retrieval, and a git-like commit history for every memory. Runs as a single local binary.

SQLiteGraphAudit trail

MCP Layer

One store shared by every assistant, over local stdio or remote HTTP. Explain yourself once.

MCPOAuth 2.1Proxy

macOS App

A native manager for your memory: browse what you know, edit blocks, walk the audit timeline. No terminal required.

MenubarGalaxy viewBlocks

Enterprise

The same engine on your infrastructure, optionally fully air-gapped, with local LLM and embeddings bundled in one signed binary.

On-premAir-gappedSLA

Memory that infers,updates and forgets.