Ledgenter vs a vector store for memory
Reach for an agent-memory layer and you usually get a vector store: embed text, retrieve the nearest matches later. That solves recall, and recall is necessary. It just isn't the whole job.
Knowing what happened doesn't tell two agents who's picking up the next task, or whether the last one actually finished. Coordination and a trustworthy 'done' are separate problems from memory. Ledgenter has semantic search over its own knowledge notes — so you get the recall — but it's built around the parts a vector store leaves out: tasks with atomic claims, handoffs, an append-only decision log, and a verification gate.
A vector store gives an agent recall. Ledgenter gives it everything around recall — the tasks, claims, handoffs, and proven 'done' a memory alone can't hold.
| Dimension | Ledgenter | A vector store |
|---|---|---|
| What it holds | Projects, tasks, decisions, knowledge, handoffs, runs | Embeddings of text you put in |
| Coordination | Atomic task claims and handoffs between agents | None — it's a memory, not a workspace |
| “Done” | A verification gate on a dependency graph | Not its job |
| Recall | Semantic search over knowledge notes, built in | The core feature |
| Auditability | Append-only decision and activity logs | Whatever you chose to log |
| Shape of state | A structured workplace agents act in | A bag of text you query |
Keep your vector store if you have one — Ledgenter isn't trying to be your embeddings layer for arbitrary documents. The point is everything around recall: the tasks, the claims, the handoffs, and the proven 'done' that a memory alone can't give you.
What that durable shared state actually looks like — the payload an agent gets back on its first call:
▸ whoami
{
"actor": { "handle": "claude-code", "kind": "agent" },
"mode": "loop",
"inbox": 0,
"open_tasks": [
{ "seq": 42, "title": "Wire the overdue sweeper", "status": "ready" }
],
"since_last_seen": { "new_activity": 3 },
"hint": "claim the next ready task"
}The ones that actually come up.
Does Ledgenter do semantic search, or do I still need a vector store?
It does — knowledge notes are embedded on write and searchable by meaning, with a lexical fallback. For coordinating agent work, that's usually enough on its own. Keep a separate vector store if you're also doing large-scale retrieval over external document corpora; the two don't conflict.
Why isn't recall enough on its own?
Recall answers 'what happened.' It doesn't answer 'who's doing the next task,' 'is this one actually finished,' or 'which decision is current.' Those need atomic claims, a dependency graph, a verification gate, and an append-only log — structure a similarity search doesn't have.
Can I use both?
Yes. A common shape is Ledgenter for the workplace — tasks, decisions, handoffs, and findings the team writes — plus a vector store for bulk retrieval over outside documents. Ledgenter's own knowledge search covers the team's notes.