Agent Memory

Agent memory lets an agent retain facts about an end user across separate conversations. When enabled, facts are extracted from the conversation after it ends and injected back into the system prompt on subsequent turns with the same user.

How It Works

1. A conversation runs normally. Memory has no effect during the first session.
2. After a session accumulates 3 or more user turns, a background extraction pass runs using the LLM. It reads the conversation and produces a list of deduplicated facts about the user (name, preferences, stated context, etc.).
3. Facts are stored in the _fyso_agent_memory table, keyed by (tenantId, agentId, externalRef) — where externalRef is the identifier your channel uses to distinguish end users (e.g. a session token or user ID).
4. On the next conversation with the same user, the 10 most recent facts are retrieved and appended to the agent's system prompt before the first LLM call.

Memory extraction is asynchronous and adds roughly one extra LLM call per qualifying conversation.

Enabling Memory

Memory is disabled by default. Enable it on the agent config:

curl -X PUT https://api.fyso.dev/api/agents-config/<agentId> \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{ "memory_enabled": true }'

Or toggle it from the agent edit page in the admin panel.

Database

Memory requires migration 0068. The migration creates the _fyso_agent_memory table:

Column	Type	Description
id	uuid	Primary key
tenant_id	text	Tenant scope
agent_id	text	Agent scope
external_ref	text	End-user identifier (from the channel)
fact	text	Extracted fact string
created_at	timestamp	When the fact was stored

Behavior Details

Condition	Result
memory_enabled: false (default)	No retrieval, no extraction. Existing agents unaffected.
Session has fewer than 3 user turns	Extraction does not run.
Session has 3+ user turns	Extraction runs after the session; facts are deduplicated before insert.
Memory retrieval at conversation start	Up to 10 most recent facts injected into system prompt.

Privacy and RGPD

Facts stored in _fyso_agent_memory contain information about end users. If you operate under GDPR / RGPD, ensure that:

• Your DPA with Fyso covers memory storage (see RGPD Compliance).
• Data suppression via DELETE /api/rgpd/users/:externalRef/ai-data also deletes memory facts for that user.
• Users who have refused AI consent (ai_consent: false) will not trigger memory extraction because the LLM is never called.

Limitations

• The in-process memory store runs on a single API container. In multi-process deployments, facts are still durable (stored in Postgres) but extraction is fire-and-forget per process.
• Fact extraction quality depends on the configured LLM and its context window. Very long conversations may result in partial extraction.
• Memory is scoped per agent. Facts from agent A are not visible to agent B even for the same user.