Concepts¶
agent-memory pairs a deterministic memory substrate with a lightweight cognitive loop. Together they turn memory-aware agent work into predictable innovation with human partnership — bold ideas, faithful delivery, a human in the loop at every altitude.
-
The backward layer: an event-sourced ledger projected into live state, with deterministic decay, supersession, and provenance. Faithful to what happened.
-
The forward layer: Current State → Vision → Blueprint → Design → Implementation → Feedback, with an enforceable intent trace. Faithful to what was intended.
-
Reusable capabilities authored once in a neutral
agent-skills/layer and run by any agent — plus seven tool-managed built-ins. -
How facts strengthen, fade, and archive — by counting session files, never by a floating-point score.
-
How the ritual actually happens without depending on an agent self-triggering: triggers, CI, init, and the judgment-vs-arithmetic boundary.
The mental model¶
There are two distinct memory layers in a developer's life, and agent-memory keeps them separate by design:
| Layer | Where it lives | Who owns it | What it holds |
|---|---|---|---|
| Personal | ~/ (home) |
You | Your vendor's chat history, model prefs, global settings |
| Shared (team) | the repo's memory/ |
The team, via git | Project rules, decisions, session logs |
The tool only ever touches the team layer. Your personal ~/.claude/, ~/.cursor/,
Application Support, AppData — none of it is read, modified, or moved. Whatever vendor you
prefer keeps working exactly as before.