Agents

Memory & sessions

An agent can remember facts across conversations, and its activity is grouped into sessions that are anchored on-chain — so a run can be pointed to and verified later.

Memory tools

Enable the memory_remember and memory_recall tools to give the agent a private memory namespace. The model decides when to save something worth keeping and when to look it back up; each agent's memory is its own, scoped to that agent.

{
  "name": "support-bot",
  "system_prompt": "Help customers. Remember preferences they tell you.",
  "model": "gpt-4o-mini",
  "bucket_ids": ["<bucket_id>"],
  "tools": ["kraterion_search", "memory_remember", "memory_recall"]
}

Sessions

When an agent has at least one attached bucket, its invocations are grouped into a session. List them to see when a session opened, how many invocations it holds, and its current status.

curl https://api.kraterion.com/v1/agents/<agent_id>/sessions \
  -H "Authorization: Bearer kr_live_..."
# → [{ id, status, opened_at, last_activity_at,
#       invocation_count, tx_digest, ... }]

On-chain anchoring

A session flushes when it goes idle and anchors its state on-chain, producing a transaction digest (tx_digest). That digest is a durable, third-party-verifiable handle to what the agent did during the session — not just a row in Kraterion's database. You can also end a session explicitly with POST /v1/agents/:id/sessions/:sessionId/end.

Replay

Because a session carries the model, the retrieved chunks (by content hash), and the tool calls it made, its record is enough to reconstruct what happened and why an answer came out the way it did. The tx_digest is the anchor you reference when you need to point at a specific run.