Prove what your agent knew, and when.
Engram gives AI agents a durable, tamper-evident record of what they saw and decided. Write fast, recall fast, and anchor to Bitcoin, so no one can quietly rewrite the past. Not even us.
MIT / Apache 2.0 · No token · No account needed to verify
Illustrative record. A live, independently verifiable record replaces this example at launch.
The Problem
Real agents. Fake memories.
Real consequences.
AI agents already execute workflows and manage real assets. But their context lives in ordinary databases that an operator, or an attacker, can edit without a trace. Feed an agent one fake memory and it will make a confident, catastrophic decision based on a lie.
Context manipulation
Unverifiable autonomy
The Solution
Three API calls. One audit trail.
Log fast, recall fast, settle on Bitcoin.
Built to drop into existing agent frameworks.
Log context
Record your agent's context and decisions through the SDK. Every record is hashed and committed the moment it is written.
Fast retrieval
The semantic V-DHT layer returns relevant records in about a tenth of a second in testing, keeping the agent's chain of thought intact.
Bitcoin settlement
Record roots are batched and anchored to Bitcoin through Babylon. From that moment, what your agent knew at time T is provable, and any later tampering is detectable by anyone.
Speed is for chatbots. Proof is
for autonomous agents.
Engram is built for workflows where every decision may be audited later: trading, payments, compliance, autonomous operations.
The Architecture
Two layers for speed.
One anchor for truth.
Engram separates fast retrieval from settlement, so agents get quick recall and Bitcoin-grade assurance without paying EVM storage fees.
Hot layer: Semantic V-DHT
A verifiable distributed hash table acts as the high-speed semantic cache for immediate recall.
Cold layer: Durable coded storage
Records are erasure-coded across independent providers: 1.6x storage overhead, and data stays recoverable after losing more than a third of nodes.
Bitcoin anchoring via Babylon · Tamper-evident audit trail
An MCP server is in development, so Engram will plug into existing agent frameworks in minutes.
By the Numbers
Measured, not promised.
to verify that a record exists
per storage proof, checked in 21 ms
data recovered across six failure and attack scenarios
storage overhead, versus 3x for standard replication
per GiB anchored when batched, down from $607.60 unbatched
Fine print: Measured on our research prototype and in a 500-node simulation. Methodology and limits in the system paper.
Open Source
Built in the open.
Looking for design partners.
The record format, the verification spec, and the code ship under MIT and Apache 2.0. Verification is free and needs no account. The proof format is being aligned with OpenTimestamps, so records stay verifiable even if Engram disappears. If you are building agents that handle critical tasks, we want to work with you.