Engram · The memory faculty

Agents that
remember.

Episodic and semantic memory for AI agents. A queryable record of every thought, every decision.

Engram is the Memory faculty of the Ginnung cognitive runtime. It captures agent experiences as structured events — observations, reasoning, actions, outcomes — and makes them retrievable by similarity, time, or causal chain. The result: agents that learn from their own history, and an audit log your compliance team can actually read.

What is Engram

Memory is a substrate,
not a feature.

Most AI memory systems are vector databases with a wrapper. They store chunks, retrieve by similarity, and call it memory. Engram is built differently.

Every interaction your agent has becomes a typed event: an engram — the trace a single experience leaves behind. Events carry their full context — the prompt, the model output, the tools called, the outcome — and the relationships between them. You query by meaning, by time, by causal chain, by agent, by outcome.

The same data structure powers two surfaces: agents read it to ground their reasoning, and humans read it to understand what their agents actually did. One source of truth. No reconciliation between “what the agent saw” and “what the audit captured.”

What it does

Capabilities

Ingest

Typed event stream

SonderEvent envelopes from any agent runtime — North, LangGraph, AutoGen, custom. Schema-validated at the edge. No untyped blobs.

Embed

Multi-model ensemble

Embed once across multiple models. Query against any of them, or all of them weighted. No vendor lock-in on the retrieval layer.

Retrieve

Semantic + structural

Vector similarity is one query shape. Time windows, causal chains, agent histories, outcome filters — all first-class.

Govern

Audit-grade by default

Append-only event log. Cryptographic chain. Replay any agent decision against the exact memory state it had at the time.

Surface

Dashboard at openengram.ai

Browse, search, and visualize agent memory. Built for engineers debugging behavior and compliance teams reading the trail.

Integrate

One SDK, one schema

TypeScript, Python, Rust clients. Same SonderEvent schema across every Ginnung faculty. Memory composes with reasoning, governance, and the rest.

The audit log is the product

August 2, 2026.

EU AI Act Article 12 takes effect August 2, 2026. High-risk AI systems must produce automatic logs of their operation — sufficient to identify risks, trace decisions, and demonstrate compliance. Penalties reach €15 million or 3% of global turnover.

Most teams will discover this six weeks before the deadline and retrofit logging onto systems that weren’t designed for it. The logs will be incomplete, the schemas inconsistent, the queries slow. Auditors will not be impressed.

Engram inverts the problem. The memory your agents use to think is the audit log. The same query that grounds a model’s next response produces the artifact regulators read. No second system. No reconciliation. No retrofit.

Get started

Two ways in.

Hosted

openengram.ai

Sign in, point your agent SDK at the ingest endpoint, get a dashboard. Free during the open beta.

Self-host

github.com/heybeaux/engram

Docker compose, your hardware, your data. Same schema as the hosted tier. MIT-licensed core.

Built by

Engram is the Memory faculty of Ginnung, the cognitive runtime built by heybeaux. The core is open-source, the SonderEvent schema is public, and the hosted dashboard lives at openengram.ai.