A vector database stores embeddings. Agent Memory decides what to keep and how to recall it. It combines vector, full-text, structured, graph and episode searches, consolidates sessions into durable facts and procedures, resolves contradictions, tracks provenance and validity time, and returns a context sized to a token budget. pgvector is one of the components it uses.

There are four. Session memory holds context for the current run in Redis. Workspace memory holds durable facts. Episodic memory keeps one record per run with its outcome. Procedural memory holds learned how-tos. Consolidation moves knowledge from the session into the other three, and memory blocks give you hand-curated context alongside them.

Use MCP if your agent is an MCP client: there is a remote Streamable HTTP endpoint with OAuth, a stdio server, and a Claude Code plugin. Otherwise use the Node or Python SDK, the .NET client, or the /v1 HTTP API. All of them share one contract, and a parity scenario in CI checks that they behave the same.

In the database. Every tenant table is protected by Postgres row-level security keyed to a transaction-local tenant setting. The application role cannot bypass RLS, and a missing tenant context returns zero rows. Composite foreign keys block cross-tenant references, and Redis keys are scoped per tenant as well. Within a tenant, workspaces separate data and per-memory visibility controls what each agent can see.

Yes. A delete redacts the memory right away and removes its derived copies in the background, and you can check the erasure status until it reads complete. If other memories derive from the one you are deleting, the delete is blocked and returns their ids, unless you ask for a cascade. Admins can also purge or export a whole workspace or tenant.

There is a reference air-gapped deployment profile that runs Postgres, Redis, a local OpenAI-compatible model server and the Agent Memory hosts with no outbound network. It has passed an acceptance run. Contact us to discuss self-hosted and on-premises deployments.

Extraction, summarization and reranking can use Anthropic Claude or any OpenAI-compatible endpoint, including local servers such as Ollama or vLLM. Embeddings can use OpenAI, AWS Bedrock or a local ONNX model. If no LLM key is configured, the platform falls back to rule-based processing.

We ran the full LongMemEval_S set (500 questions) with GPT-4o as the answerer and the official evaluate_qa.py script with the GPT-4o judge, and got 83.8% QA accuracy (95% CI 80.6–87.0%). An earlier 50-question subset scored about 4 points higher, so we publish the full-set number. Methodology and the deviations from defaults are documented with the results.