In developmentEngram is being built in the open — the first public release is coming.Get notified →
engram.is
A memory thatcompounds
Engram stores what you and your agent learn as small, linked, provenance-sealed
pages — and autosearches the relevant ones back into every session. Knowledge compounds
instead of being re-derived. The engine is open; your store never leaves your box.
$ engram ask "how did we handle the rate-limit retries"[[retry-backoff-policy]] · traced · used 4[[idempotency-keys-gotcha]] · traced(2 pages · 1,812 bytes read)
See it work
Four things Engram does. Watch each one run — then read what it means.
You ask in plain language — no keywords, no query syntax. Engram reads what
you're working on, walks its index, and opens only the pages that actually answer, each one
cited with where it came from. The thing your agent figured out last week is simply
there in this session, without you remembering it exists.
The same question through plain grep drags back every file that
mentions the words — tens of thousands of tokens of haystack to wade through. Engram reads one
index, opens the three pages that answer, and stops. On a real store that's about
86% fewer bytes than grep -rin prints — measured from the real command and
printed, never a number we ask you to trust. Fewer tokens, less waiting, an answer instead of
a pile of matches.
A page's trust is never a score someone typed in — it's worked out fresh on
every read from the evidence underneath it: traced or only inferred, how recent, how well
corroborated. It's symbolic — explicit rules you can inspect, nothing learned, never
popularity — an approach inherited from Titan and Mnemosyne, the production systems behind
Engram. (Separately, an optional embedding pass can suggest latent links between pages
— a curation aid that never touches recall and never overrides the evidence.)
Every page is hash-chained to the one before it, so the whole memory is a
single tamper-evident line — alter one word in a month-old page and the chain visibly breaks.
Now and then the tip of that chain is anchored into Bitcoin, timestamping your knowledge
against the most secure clock there is. It proves when a thing was recorded and that
it hasn't moved since — provenance, never veracity.
▸
The thesis
Engram is opinionated about honesty. Four ideas do the work.
Provenance, not veracity
A sealed record proves a thing was recorded at a time, unaltered — never that it is
true. Every page carries an evidence state (traced / inferred
/ stale), and Engram surfaces it. Nothing inferred is ever dressed up as fact.
A proof that can die
Trust is earned by falsification: every claim ships with a test that is sabotaged
to prove it can fail. A proof that survives its own sabotage is a decoration. What can die is
worth something; what merely passed is not.
Recall beats grep
Retrieval is index-first: read one catalog, pull only the pages that answer
the query, cite them. On a real store, broad questions read ~86% fewer bytes
than what grep -rin prints — measured, net of cost, never a black-box percent.
Decay by evidence
Freshness and trust are never a stored score — they are derived from the
record every time a page is read. A trust value can't rot while the evidence under it moves.
Usage never inflates rank; only evidence does.
How it works
1
Pages, linked
Systems, frames, gotchas,
decisions — one durable thing per page, cross-linked like a wiki.
2
Sealed to a chain
Every page is
hash-chained (optionally anchored to Bitcoin), so your memory is provably, immutably continuous
across sessions, machines, and weeks.
3
Autosearched in
A session-start hook
surfaces the relevant pages automatically; a session-end hook digests the conversation into a
new page. It runs itself — you never invoke it.
4
Everywhere you code
One MCP server.
Verified in Claude Code; protocol-compatible with Codex, Gemini, Cursor,
Cline, Windsurf, Zed and Continue — they speak MCP, but we have not verified them ourselves.
See editors.
Encrypted before it leaves your machine
Engram encrypts your store on this machine, before anything leaves it. That means we cannot read it — and we cannot recover it. If you lose your recovery phrase and no enrolled device remains, your store is gone permanently. No support request can undo that. Keep the phrase somewhere safe and offline.
That is the trade Engram makes on purpose: we cannot read your thinking, and the
price of that is that we cannot rescue it either.
Then, once, in a terminal — the bootstrap installs the engine on your PATH:
# macOS / Linux
bash setup.sh
# Windows
powershell -ExecutionPolicy Bypass -File setup.ps1
…and /reload-plugins. Your store lives in ~/.engram/store,
owner-only, and is never pushed anywhere.
On the roadmap
Built in the open, one honest claim at a time. No dates — each piece ships when it's proven, not before.
Now
Index-first recall, trust you can audit, the sealed hash-chain (optionally anchored to
Bitcoin), and one MCP server that plugs into every editor. Cross-project memory, today.
Next — first public release
Useful on day one — it seeds itself from the notes you already have
(CLAUDE.md, MEMORY.md). A memory for a single project,
not just across them. A lightweight essentials protocol so your agent actively
saves and links what it learns — plus proactive surfacing that offers the right
page at the right moment. Mac & Windows installers.
Later — next major
Memory that syncs across your machines, encrypted, your keys. And the advisor
grows sharper — semantic surfacing and a feedback loop that learns what to surface — claimed only
once we can measure it does. No bold claims before the number exists.
Always
More terminals (Codex, Cursor, Cline, Zed, and the rest), steady hardening, and a direction
shaped by what you actually tell us it needs.
Get updates
One short note when Engram ships a release worth your time. No noise, unsubscribe in a click.