How it works

Three verbs. One graph that grows over time.

memnode gives your agent a place to write things down, read them back later, and correct them when they change. That's the product. Below is what each step actually looks like.

Record

Your agent writes down something it just learned — a user preference, a fact, a decision.

record_memory({
  type: "observation",
  summary: "Alice prefers SQLite for prototypes",
  source: "user",
  epistemic_type: "reported"
})

Recall

Later — next session, next week — it asks a question and gets back what it already knows.

recall_memory({
  question: "What database does Alice prefer for prototypes?"
})
// -> "SQLite (reported by user, 2026-02-14)"

Correct

When something changes or turns out wrong, the old memory is superseded — not silently overwritten.

correct_memory({
  supersedes: 42,
  summary: "Alice now prefers Postgres for prototypes",
  reason: "switched after the Jan outage"
})

Anatomy

What a memory actually is

Not a blob of text. Each memory has a type, a source, a claim about how the agent knows it, and edges to related memories. That's what lets recall be precise — and correction be traceable.

summaryAlice prefers SQLite for prototypes

typeobservation

epistemic_typereported

sourceuser · 2026-02-14

edgesabout → "Alice" · relates_to → "prototype setup"

statusactive (not superseded)

What happens over time

Your agent starts remembering you

Day 1

You add a handful of memories manually. Your agent immediately has context it would otherwise forget.

Day 14

Dozens of memories. Recall starts surfacing things you told it once, weeks ago, without prompting.

Day 90

Wrong facts get corrected, not deleted — the lineage stays so you can audit what the agent believed and when.

Day 365

Unused memories fade; reinforced ones stay. The graph is now a real knowledge substrate, not a log file.

Ready to plug it in?

The quickstart covers the three verbs end-to-end in about five minutes.

Read the quickstart Create a workspace