Agents that remember and improve

How it works

For the person retyping the same context every session

If you correct the same misunderstanding on every run — the house style, the fact that you use Postgres and not MySQL, the client who must never be emailed after 5pm — that correction is the work. Memory is where it goes so it only has to be said once.

1. Install AGNT and open it. The local API it talks to runs at http://localhost:3333/api/.
2. Pick the agent that does the recurring job. Memory is stored per agent, so the choice matters.
3. Let the agent record what it learns as it works, or write entries yourself. Each entry is a memoryType of fact, preference or correction, plus its content.
4. Entries carry a relevance_score and are retrieved when they are relevant to the task at hand, rather than all being replayed into every prompt.
5. Recall across your wider history — past conversations, runs, outputs and insights — is a separate keyword search over the same local database.

You, once, in an ordinary sentence:

  "We only deploy on Fridays after the staging soak passes."

Recorded as a single typed entry:

  POST /api/insights/memory/:agentId
  {
    "memoryType": "preference",
    "content": "Deploys ship Friday only after the staging soak passes."
  }

A later run, days afterwards, reading it back:

  GET /api/insights/memory
  {
    "success": true,
    "count": 1,
    "memories": [
      {
        "memory_type": "preference",
        "content": "Deploys ship Friday only after the staging soak passes.",
        "relevance_score": 0.9
      }
    ]
  }

...so the agent proposes the Friday window and says which
memory it is acting on, instead of asking you again.

• Memory is per agent, not a shared team brain. Two agents do not automatically know what the other learned.
• Delete an agent and its memories are orphaned rather than silently swept up. Cleanup is an explicit call (DELETE /api/insights/memory/orphaned), and the orchestrator’s own memory is always preserved.
• Keyword recall sanitises queries to letters, digits, - and _ and prefix-matches them, so deploy also finds deployment. It is deliberately forgiving, which makes it a poor tool for exact-phrase or punctuation-sensitive search.
• Insights the system derives from your runs are pending by default — recommendation-first. Nothing changes an agent until it is applied.
• Entries are stored locally and scoped to your user, with no cross-user visibility. That is not the same as nothing leaving your machine: if a step calls a hosted model, the content of that step still goes to that provider. Point the step at a local model if that matters.

Next: the full Memory & Insights API reference, or Getting Started if you have not installed AGNT yet.