Uploaded October 2026 | Updated November 2026, 1 day ago
Read more here: blogs.oracle.com/developers/from-prompt-to-persistence-part-1-designing-multi-tenant-agent-memory-schemas-for-saas?source=:ad:vd:yt:awr:a_apac::W4yt&SC=:ad:vd:yt:awr:a_apac::W4yt
How do you secure multi-tenant agent memory? In this Oracle Developers Explains video, Jeremy Daly shows how tenant identity becomes a database rule: Oracle Virtual Private Database row-level security in Oracle AI Database 26ai enforces the tenant boundary on every read, write and delete, so a forgotten WHERE clause, a retrieval bug or a prompt that asks for the wrong customer's data cannot reach another tenant's memory.
00:00 Same query, different tenant
00:43 What this video covers
01:10 The tenant is the access boundary
02:14 Row-level security with Oracle Virtual Private Database
03:03 Four scope columns, one enforced boundary
03:37 Shared memory is a scope, not a separate table
04:42 One retrieval query across every memory type
06:30 The memory manager: one controlled path
07:15 Versioning: supersede, don't overwrite
07:45 Right to forget: deletion through provenance
09:03 Set once in the session, enforced by the database
In the cold open, the same queries run against the same tables for two SaaS customers, Acme and Globex. Globex sees none of Acme's rows, and no tenant ID filter is written in the query: the tenant identity attached to the database session decides what comes back.
Four principles follow from the tenant boundary. Apply the same boundary to every memory type. Share memory by widening its scope. Version every change to preserve history. Use provenance to delete source and derived data. Oracle Virtual Private Database attaches one tenant policy to every tenant-scoped memory table, covering selects and every mutation, and the update check stops an Acme request from inserting or moving data into Globex.
Every durable memory row carries four scope columns: tenant ID, user ID, agent ID and thread ID. Tenant ID is required and enforced by row-level security. A null user, agent or thread ID widens a record's scope inside the tenant, so shared memory is a scope, not a separate table: a refund guideline written once for all of Acme is inherited by every Acme user.
Guidelines, personas and tools are enumerated, entities and summaries use hybrid search, workflows and knowledge base chunks use vector similarity, and recent conversation events return as an ordered tail. In Oracle AI Database 26ai one retrieval query combines every memory type with UNION ALL under the same tenant policy and transactional snapshot, and vector distance is computed only over rows the tenant policy already let through.
The memory manager sets the session identity on every connection, runs writes as transactions that require provenance, and traces deletion through source links. A changed fact becomes version 2 and closes version 1 instead of being overwritten. A right-to-forget request follows provenance to tenant-wide facts derived from the user's conversation, and one transaction scrubs derived content and embeddings, deletes the events and writes an audit record.
Row-level security, from 02:14, enforces the tenant predicate even when a query omits it; this Oracle Developers article applies the same Virtual Private Database pattern to governing agent memory:
medium.com/oracledevs/using-jev-and-oracle-ai-database-to-govern-agent-memory-65365670fa5c
The scope predicate, from 03:37, handles users, agents and threads while row-level security handles tenants; the final article in the series builds context assembly on top of it:
medium.com/oracledevs/context-assembly-for-multi-tenant-agent-memory-budgets-selection-and-tool-calls-with-mcp-4e99dab11cf4
Read Jeremy Daly's two-part series, From Prompt to Persistence. Part 1 designs the multi-tenant memory schema:
blogs.oracle.com/developers/from-prompt-to-persistence-part-1-designing-multi-tenant-agent-memory-schemas-for-saas
Part 2 puts the schema to work, with the retrieval query, versioned supersession and the right-to-forget cascade:
blogs.oracle.com/developers/from-prompt-to-persistence-part-2-putting-the-multi-tenant-agent-memory-schema-to-work
Run every database operation from this video in the companion notebook:
github.com/oracle-devrel/oracle-ai-developer-hub/blob/main/notebooks/multitenant_schema_walkthrough.ipynb
Subscribe to Oracle Developers for more on building AI agents.
#AIAgents #AgentMemory #OracleAI
Read more here: blogs.oracle.com/developers/from-prompt-to-persistence-part-1-designing-multi-tenant-agent-memory-schemas-for-saas?source=:ad:vd:yt:awr:a_apac::W4yt&SC=:ad:vd:yt:awr:a_apac::W4yt
How do you secure multi-tenant agent memory? In this Oracle Developers Explains video, Jeremy Daly shows how tenant identity becomes a database rule: Oracle Virtual Private Database row-level security in Oracle AI Database 26ai enforces the tenant boundary on every read, write and delete, so a forgotten WHERE clause, a retrieval bug or a prompt that asks for the wrong customer's data cannot reach another tenant's memory.
00:00 Same query, different tenant
00:43 What this video covers
01:10 The tenant is the access boundary
02:14 Row-level security with Oracle Virtual Private Database
03:03 Four scope columns, one enforced boundary
03:37 Shared memory is a scope, not a separate table
04:42 One retrieval query across every memory type
06:30 The memory manager: one controlled path
07:15 Versioning: supersede, don't overwrite
07:45 Right to forget: deletion through provenance
09:03 Set once in the session, enforced by the database
In the cold open, the same queries run against the same tables for two SaaS customers, Acme and Globex. Globex sees none of Acme's rows, and no tenant ID filter is written in the query: the tenant identity attached to the database session decides what comes back.
Four principles follow from the tenant boundary. Apply the same boundary to every memory type. Share memory by widening its scope. Version every change to preserve history. Use provenance to delete source and derived data. Oracle Virtual Private Database attaches one tenant policy to every tenant-scoped memory table, covering selects and every mutation, and the update check stops an Acme request from inserting or moving data into Globex.
Every durable memory row carries four scope columns: tenant ID, user ID, agent ID and thread ID. Tenant ID is required and enforced by row-level security. A null user, agent or thread ID widens a record's scope inside the tenant, so shared memory is a scope, not a separate table: a refund guideline written once for all of Acme is inherited by every Acme user.
Guidelines, personas and tools are enumerated, entities and summaries use hybrid search, workflows and knowledge base chunks use vector similarity, and recent conversation events return as an ordered tail. In Oracle AI Database 26ai one retrieval query combines every memory type with UNION ALL under the same tenant policy and transactional snapshot, and vector distance is computed only over rows the tenant policy already let through.
The memory manager sets the session identity on every connection, runs writes as transactions that require provenance, and traces deletion through source links. A changed fact becomes version 2 and closes version 1 instead of being overwritten. A right-to-forget request follows provenance to tenant-wide facts derived from the user's conversation, and one transaction scrubs derived content and embeddings, deletes the events and writes an audit record.
Row-level security, from 02:14, enforces the tenant predicate even when a query omits it; this Oracle Developers article applies the same Virtual Private Database pattern to governing agent memory:
medium.com/oracledevs/using-jev-and-oracle-ai-database-to-govern-agent-memory-65365670fa5c
The scope predicate, from 03:37, handles users, agents and threads while row-level security handles tenants; the final article in the series builds context assembly on top of it:
medium.com/oracledevs/context-assembly-for-multi-tenant-agent-memory-budgets-selection-and-tool-calls-with-mcp-4e99dab11cf4
Read Jeremy Daly's two-part series, From Prompt to Persistence. Part 1 designs the multi-tenant memory schema:
blogs.oracle.com/developers/from-prompt-to-persistence-part-1-designing-multi-tenant-agent-memory-schemas-for-saas
Part 2 puts the schema to work, with the retrieval query, versioned supersession and the right-to-forget cascade:
blogs.oracle.com/developers/from-prompt-to-persistence-part-2-putting-the-multi-tenant-agent-memory-schema-to-work
Run every database operation from this video in the companion notebook:
github.com/oracle-devrel/oracle-ai-developer-hub/blob/main/notebooks/multitenant_schema_walkthrough.ipynb
Subscribe to Oracle Developers for more on building AI agents.
#AIAgents #AgentMemory #OracleAI



