Provenance, gaps & capture

Where memories come from

Every memory records the surface it was formed on — the conversation, code session, or capture context, and the task within it — and an audit spine of action records ties messages, articles, and memories together as events: created, updated, attached, informed. Source discriminators note which tool a conversation came from and which repository or file an article traces to. The result is a durable trail from any memory back to its origin, which is what makes the belief revision on the previous page auditable rather than opaque.

Alongside where, ENGRAM records by which credential. A person or an agent can hold several access tokens, and until recently they were interchangeable in the record — so the coding agent learned this while working in the repo and the desktop assistant was told this by a document it ingested looked identical after the fact. They carry very different evidential weight. Each memory and article now carries the token that wrote it, on the record itself rather than only on the audit spine, which matters because most memories have no separate audit event to hang it from.

Retire, don't erase

Delete an article or document and its imprinted memory isn't destroyed — it's retired: marked with the moment it stopped being valid and a reason, then excluded from recall. From a reader's point of view it's gone instantly, byte-for-byte identical to deletion. But the belief is preserved for audit and for any future supersession chain, so history never quietly disappears.

Capture across your tools

Research doesn't only happen in one place. ENGRAM can register finished conversations from your other assistants and file them into the same graph, idempotently, so re-capturing the same session appends rather than duplicates. Your knowledge base grows wherever you work.

External agents drive capture — and everything else — through a Model Context Protocol server; see Connecting External Tools. The story of tracking provenance and grouping by project is on the blog.

A mind that knows what it's missing

ENGRAM doesn't only accumulate; it audits its own coverage. It surfaces detected gaps — a topic that's thin, a missing bridge between clusters, an alias that should be merged — and can draft a proposed closure: an article, a bridge, or an alias fix, held for human approval before anything is executed. Nothing is patched into the graph behind your back.

More on cross-source capture and self-closing gaps: Completing the Mind and A Mind That Knows What It's Missing.