The Advisor Loop

Maturity is not something you wait for — it is produced by a four-step cycle run on every task. Store what was decided, link it to what it relates to, let engrams surface it at the right moment, and act on better information. The outcome of step four is the input to step one. The loop compounds.

1 Store the agent logs decisions, patterns, progress as work happens
2 Link connect them with refines, extends, depends_on
3 Surface prime, relevant and graph bring it back at the right moment
4 Act you make a better-informed choice — which gets stored

…and step 4 feeds straight back into step 1.

1 · Store — capture reasoning at the moment it exists

The agent logs while the context is hot, not at session end: engrams decision log --summary "…" --rationale "…" for choices, engrams pattern log for conventions worth enforcing, engrams progress log for the work itself. Rationale is the payload — a decision without its why is just a fact; with it, it is advice waiting to happen.

2 · Link — turn records into reasoning

engrams link add --rel … connects what was just stored to what it builds on. The nine relationship types — depends_on, extends, implements, part_of, refines, relates_to, supersedes, supports, uses — are a validated ontology, not free text: disjoint pairs, domain/range constraints, and cardinality are rejected at link time unless you pass --force. Links are where the graph stops being a pile of notes and starts being an argument.

3 · Surface — the right context, unprompted

This is the step that makes engrams an advisor rather than an archive. engrams prime loads the decisions, patterns, and active work relevant to the session before anything else runs. engrams relevant <paths> pulls the reasoning anchored to the exact files about to be edited. engrams graph and engrams decision search answer the exploratory questions. Nothing here waits for you to remember that the context exists.

4 · Act — decide with the full picture, then feed the loop

You and the agent act on advice instead of guesses: a settled debate stays settled, a registered pattern passes engrams check --staged before it can be violated, a contradiction gets caught while it is still a diff. And whatever you decide next is a new decision — stored, linked, and ready to advise the next session. The output of the loop is a smarter loop.

Compounding, measured This repo ran the loop for 26 days: one link on day 1 became a 29-link reasoning web — amplified by engrams into 802 total edges. Nothing about the tool changed in that time. The graph did. That is the entire thesis.

Run the loop

The loop has two operators. You run the surfacing and enforcement side — see For You. Your agent runs the storage and linking side — see For Your Agent. For why density makes each turn smarter, read How Memory Matures.