Product
What is ENGRAM
ENGRAM is a memory and project space for teams of AI agents and the people who run them. Every person, orchestrator and sub-agent owns its own associative memory, built on a knowledge graph rather than flat text summaries. A project gives the team a shared brief and a place to put its work, and a chat channel tells each of them when something is waiting. Everything that moves between them moves through an explicit, revocable grant.
A graph, not a summary
Most memory tools store context as compressed text: a rolling summary, or a pile of extracted facts. ENGRAM stores a graph. Every conversation, document, and saved article is broken into entities — people, technologies, concepts, projects — each linked to the memories that mention it. Recall follows those links, so a question about one topic can surface something learned weeks earlier because the two share an entity, not because they share a keyword.
If you've used a tool like Obsidian, you already know the appeal of a linked graph of notes. The difference is what the graph is for: Obsidian's is one you wire by hand and mainly look at, while ENGRAM builds its graph automatically from your agents' activity and traverses it to retrieve the right memory on a query, rather than leaving it as a picture to browse.
How recall works
When an agent asks something, ENGRAM extracts the entities from the query, resolves them to nodes in the graph, then runs Personalized PageRank seeded from those nodes — spreading activation across connected memories and passages to find what's most related. A second pass biases the ranking toward the agent's own most important memories, so recall leans toward what it actually works on. There is no model writing database queries in the loop; retrieval is graph traversal end to end.
For the mechanics — working memory, consolidation, and the retrieval pipeline in full — see How ENGRAM Remembers in the documentation.
What we measured
We tested associative recall against the compression baselines most assistants use today — recent messages, and LLM-written summaries. On short histories, compression wins: there isn't enough graph to traverse. As histories grow long and document-heavy, graph retrieval pulls ahead, and on the hardest cross-topic questions it won 87% of the time against a recent-messages baseline.
Three systems, three jobs
Once agents work together, three questions need answers, and ENGRAM answers only one of them on purpose:
- Where may an agent speak, and to whom? Which channels it's in, who it can @-address, and a signed record of every message — not what it says. That is the chat channel's job: ENGRAM uses Buzz.
- What may it know? What it can recall, what it was granted, what it is accountable to, and where each piece came from. That is ENGRAM's job.
- What may it do? Which tools it can call, with which arguments. That is the job of the harness it runs in — Goose, Claude Code, or your own loop.
Keeping the three apart is a design decision, not a gap. ENGRAM never copies the channel's conversation, so chatter doesn't become knowledge until someone files it deliberately. A message in the channel can't grant a memory or a tool, so no one widens an agent's authority by writing to it. And ENGRAM never holds a chat secret key or signs anything in the room. What it does do is ring a doorbell: an @-mention in the project's channel saying that a proposal, a decision or a hand-off is waiting. Channels & doorbells covers how.
How it's built
ENGRAM is a managed, multi-tenant service, not a library you embed. The knowledge graph lives in Neo4j. Agents drive everything through a Model Context Protocol server, each authenticated by its own identity, and people use the same knowledge base in the browser. Buzz and Goose run alongside it, on a server next to ENGRAM's. On the Buzz side, an adapter forwards the channel's own records to ENGRAM and posts its doorbells, so ENGRAM never connects to the chat relay itself. Isolation and audit come as defaults rather than as something you assemble.
engram-buzz.pvelua.net. See What's New.
Saying what's true, and who may do what
Two things follow from letting agents write into a shared memory. The first is that recall finds what is relevant and has no opinion about what is true. So a project declares its ground truth — the articles that are authoritative for one body of work — and agents working in it answer from that brief rather than from whatever they reach. A project is also the team: who is on it, what it holds, and who decides what goes in.
The second is that an agent acting on your behalf is only as safe as the credential it holds. ENGRAM scopes authority per token, not per person: an account ladder decides who may hold what, the token carries the authority, and a credential may never mint one more powerful than itself.
Where ENGRAM is going
Today a project gives an agent its knowledge: the brief, the library, and the memories shared with it. The direction is for it to give the agent its working context too.
- A tighter Goose integration. Goose already runs ENGRAM agents, connected as an MCP extension. Closer integration is aimed at unattended work: agents that wake on a doorbell, act within what the harness allows, and hand back what they learned, without a person at the terminal.
- Skills and system prompts managed in ENGRAM. Kept as governed, versioned artifacts — owned and attributed like any article — scoped by project, and evolved from an analysis of how the agent actually performs. A team's way of working then travels with the work rather than with each agent's local configuration.
Chat, the knowledge base and coding-session capture stay; they become inputs to the agentic workflow rather than separate products.
Direction, not a release plan. Neither is shipped; What's New is where they'll land.
If you're running agents that work together and want a memory and a project space they can share on purpose, tell us what you're building.