What is an Advisor?
A database waits for you to ask the right question. An advisor builds enough context to raise its hand first — to tell you why a choice was made before you re-litigate it, to enforce a standard before the violation lands, to flag a contradiction while it is still cheap to fix. Engrams is the second kind of tool.
Retrieval vs. advice
| Retrieval (a database) | Advice (engrams) |
|---|---|
| Sits idle until queried. | Surfaces context at session start, unprompted (engrams prime). |
| Answers exactly what you asked — nothing more. | Answers what you should have asked: the decision behind the file you're about to edit (engrams relevant). |
| Stores records; meaning is your problem. | Stores reasoning: decisions linked by refines, extends, supersedes, depends_on. |
| Lets you repeat a settled debate. | Enforces registered patterns against your staged code (engrams check --staged). |
The difference is not a better query engine. It is that engrams accumulates a model of why your project looks the way it does — and puts that model in front of you and your agent at the moment it matters.
What engrams advises on
Recalled decisions
On 2026-08-05 this repo logged decision #44: “Path A ratified: TTSR-export-led
policy engine.” The same day, the implementation progress entry was
linked implements → #44, and the workspace-isolation bug fix (#45) was
linked part_of it. Anyone touching the policy engine now inherits the
reasoning, not just the code.
Enforced patterns
“Borrowed SQL Parameters Only” is a registered system pattern born of the
v0.8.0 zero-allocation hardening — machine-checkable, not prose in a wiki.
engrams check scans staged files against patterns like it and exits 1 on
violation. The standard is enforced before the commit, not after the review.
Caught contradictions
Two days after “Support both JSON and human-readable plain text outputs”
(2026-07-10), the project reversed course: “JSON-only CLI output” was logged
and linked supersedes to the original. The old decision stops advising but
stays auditable — settled questions stay settled.
This is engrams' own memory of building itself — explore it.
26 days of real feature work on this repository: 33 decisions, 6 patterns, 32 progress entries, connected by 29 reasoning links. Orbit it, zoom in, click a node to read what was decided and why. This is the raw material an advisor reasons over.
Node size reflects link degree. The sparse early cluster on the left and the dense ontology/policy-engine web on the right are the same project, 26 days apart — see How Memory Matures.
Advice compounds
On day 1 engrams can only fetch — the graph is nearly empty, so there is little to advise with. Every decision logged, every pattern registered, every link added makes the next piece of advice sharper. That is the compounding: the tool you run in month two is strictly smarter than the one you installed, because you taught it.