Agent memory integrations

Wire memory at the action boundary.

ContextDB connects to agent frameworks, voice stacks, support tools, MCP clients, and data sources. The shared pattern loads context, evaluates evidence before consequential tools, leaves authorization and execution to the host, and reports the result.

recall -> remember -> evaluate_action -> confirm if required -> re-evaluate -> host action -> report_execution

The Python client method evaluate_action calls POST /v1/recall_for_action. HTTP and MCP keep the recall_for_action name. MCP has no gate tool.

Runnable starter kits

Run the boundary before adapting it.

These public starters put the memory lifecycle beside the framework call that needs it. The proof column states whether CI checks a framework import, a TypeScript contract, or an offline example suite.

Framework starters

Integration	Runtime	Boundary	Proof	Status	Links
OpenAI Agents SDK	Python	`recall` before `Runner.run`. Selected writes go through a typed tool.	Framework import in CI	Runnable starter	Starter kit
LangGraph	Python	`recall` inside a graph node before model invocation	Framework import in CI	Runnable starter	Starter kit
Vercel AI SDK	TypeScript	`recall` before `generateText`. The memory tool stays server-side.	TypeScript contract in CI	Runnable starter	Starter kit
LiveKit Agents	Python / voice	Starter covers conversational `recall` and sourced writes. The guide adds host-owned action evaluation.	CI import check and 7-test offline simulation	Runnable and tested	Starter kit Example repository LiveKit Agents memory guide
PyAI Omni	TypeScript / Python / voice	Starter covers conversational recall and sourced writes. The guide adds host-owned action evaluation.	CI contract check and 10-test offline simulation	Runnable and tested	Starter kit Example repository PyAI voice agent memory guide

Browse every starter path

Integration paths

Match each runtime to its action boundary.

Framework and vendor names identify code paths and published patterns. They do not indicate a vendor partnership or endorsement.

Voice

Integration	Runtime	Boundary	Proof	Status	Links
Retell AI	Custom functions / webhooks	Gate in-call actions. Write selected facts after the call.	Published integration pattern	Integration guide	Retell AI memory guide
Pipecat	Python / voice	Example covers conversational recall and sourced writes. The guide adds host-owned action evaluation.	8-test offline simulation	Runnable and tested	Pipecat voice agent memory guide Example repository

Support

Integration	Runtime	Boundary	Proof	Status	Links
Intercom Fin	Support actions	`recall_for_action` before credits, plan changes, or account updates	Published integration pattern	Integration guide	Intercom Fin customer memory guide
Custom support agents	Any server runtime	Recall for continuity. Evaluate remembered evidence before a mutation.	Host-enforcement pattern	Integration pattern	AI support agent memory pattern
Server-only web chat	TypeScript / Node	Browser calls a server that keeps ContextDB credentials private	9-test demo-mode suite	Runnable and tested	Example repository

Protocols, clients, and lifecycle

Integration	Runtime	Boundary	Proof	Status	Links
Model Context Protocol	MCP client	`remember`, `recall`, `recall_for_action`, `pending_confirmations`, `confirm`, scoped `forget`	Published `POST /mcp` contract	MCP tools-only	MCP guide
Python and TypeScript clients	Python / TypeScript	Hosted and local clients use the documented call shapes	API and client reference	SDK documentation	API docs
Post-call Formation	HTTPS / async job	Structured turns become candidates after the realtime session	Durable submit and poll contract	Hosted Alpha guide	Formation guide
Memory CI	CLI / GitHub Action / API	Run deterministic recall and action cases against a pinned baseline	Public package and Action: Available	Hosted execution: Hosted Alpha	Memory CI guide `contextdb-memory-ci==0.1.0a1` `memory-ci-action-v0.1.0`

Data stack

MANAGED SOURCES PRIVATE ALPHA Postgres operated proof is bounded. BigQuery bounded synthetic proof used a temporary provider table. Snowflake/Databricks contract tests only cover the reader boundary. No provider has customer warehouse proof or public self-serve access.

Integration	Runtime	Boundary	Proof	Status	Links
Postgres Managed Source	Postgres protocol	`preview -> backfill -> checkpoint`	Postgres: bounded operated proof, with no customer warehouse proof.	Managed Sources Private Alpha	Managed Sources
BigQuery Managed Source	BigQuery API	Dry run byte gate before one bounded query	BigQuery: bounded synthetic proof. No customer warehouse proof.	Managed Sources Private Alpha	Managed Sources
Snowflake and Databricks Managed Sources	Warehouse SQL APIs	Bounded readers with explicit result and cursor contracts	Snowflake and Databricks: contract tests only. No customer warehouse proof.	Managed Sources Private Alpha	Managed Sources Snowflake and Databricks agent memory guide
Postgres, Supabase, and Hasura	Application database	Keep authoritative records in place. Evaluate memory before the host mutation.	System-of-record integration pattern	Guide / pattern	Postgres and Supabase agent memory guide

Run the client action flow.

Start with one framework path, one customer partition, and one host-enforced action.

Wire one framework path