How Memory Matures

Engrams does not ship smart — it becomes smart, on your project, at the speed you log to it. The data below is not illustrative. It is this repository's own knowledge graph, measured at three points over its first 26 days of real feature work.

23/1
Day 1 · nodes/links
43/11
Week 1 · nodes/links
71/29
Month 1 · nodes/links

The three stages

StageGraphWhat engrams can do
Day 1
2026-07-10		 23 nodes, 1 link Capture only. Decisions are on record — “Use SQLite context database” is the very first node — but with a single link there is no reasoning to traverse. Engrams can fetch what you logged. It cannot yet advise.
Week 1 43 nodes, 11 links Patterns emerge. The first reasoning links appear: a CLI-output reversal linked supersedes, worktree-aware resolution linked extends to workspace auto-detection. engrams relevant now returns provenance, not just records.
Month 1
2026-08-05		 71 nodes, 29 links A reasoning web. Eight ontology decisions (#21–28) all refines the schema-v3 node; the policy-engine cluster (#44, #45) hangs off the omp-integration decision by refines / implements / part_of. Chains of depends_on and supersedes let engrams answer why, not just what.

Why density is what makes it advice

Retrieval quality scales with records. Advice quality scales with connections:

  • More anchors for priming. Every linked decision and pattern is surface area for engrams prime and engrams relevant to catch on. A file you're about to edit is no longer isolated — it inherits the reasoning of everything anchored to it and everything those nodes link to.
  • More contradictions to catch. A supersedes link is only possible when both decisions exist. Density is what lets engrams check and the ontology's disjointness rules detect that your new direction conflicts with a settled one — while it is still a staged diff, not a regression.
  • Longer chains of why. Month 1's web means a question like “why does the policy engine write through to the workspace DB?” resolves through part_of → #44 → refines → the omp-integration decision. Sparse graphs answer questions; dense graphs answer follow-up questions.
The multiplier The 29 manual reasoning links are only the seed. Engrams materializes derived edges from them — inverses, transitive closures, git co-change topology. The full graph behind these 71 knowledge nodes carries 802 edges: 45 manual, 757 derived, across 67 code nodes. Every link you add is amplified.

The arc is the point

None of this required curation sessions or documentation sprints. The agent logged decisions as they were made and linked them as relationships became clear — and 26 days later the graph had crossed from storage into advice. The cycle that produces this density is short enough to run on every task: The Advisor Loop. To see the graph itself, go back to What is an Advisor?