Capabilities · Memory

Memory — the moat

Memory is the durable store that lets a run remember. It graduates a run's lessons into recallable facts, so the next task starts warmer than the last — the compounding advantage Agentry is built around.

Diagram of Memory: a durable fact and episode store backed by node:sqlite, accessed as a local stdio MCP, with a project root and a global root.
A durable fact/episode store — a local stdio MCP over node:sqlite, with project + global roots.

Memory is a durable fact/episode store: it keeps both the semantic facts a project accumulates (conventions, gotchas, decisions) and the episodes a run produces, with the provenance that earned them. It is exposed as a local stdio MCP over node:sqlite — no network service, no external database; it runs on your machine and stores on your disk.

Recall is rooted in two scopes. A project root holds what is true for this repository; a global root holds what is true across all your work. A task pulls from both, so a lesson learned once is available everywhere it applies.

This is the moat. Capability you can buy; a store of what your work has already taught the system, you can only accumulate — and it compounds with every run.

Browse the store

The Workbench includes a read-only Memory browser — search and filter the facts and episodes the store has accumulated, each with its confidence, provenance, and the reason it was kept:

The Workbench Memory browser: a master-detail view with totals (facts and episodes), a search box and type filters, a list of memory cards, and a detail panel showing a fact's confidence, creation date, and why-it-matters note.
The Memory browser — search the durable store; each fact carries its confidence, provenance, and why it was kept.