Skip to main content

​
Overview

pipecat-memcode gives a Pipecat voice agent personal, cross-session memory backed by Memcode. One MemcodeMemoryService owns two coordinated processors:
  • recall_processor() runs after the user context aggregator and before the application LLM. It searches under a strict latency budget and injects a bounded, reference-only context block.
  • capture_processor() runs after the assistant context aggregator. It stores only the finalized user/assistant pair in a tracked background task.
Recall and capture both fail open, so a memory timeout, authentication error, or service failure never prevents the voice response.

Source Repository

Package source, foundational example, changelog, and issues

Memcode Guide

OAuth setup, pipeline placement, and runtime behavior

​
Installation

The first release targets Python 3.11+ and Pipecat 1.10.x.

​
Authentication

Connect each signed-in participant to Memcode with Authorization Code, S256 PKCE, and dynamic client registration. Register with the stable public software_id ai.pipecat.memcode so Memcode can assign server-owned Pipecat analytics attribution. The identifier is not a credential or proof of installation. Registration happens once per deployment; authorization is completed once per Memcode account. Store rotating refresh tokens in an encrypted server-side store, replace each token set atomically, and implement the SDK token store’s refresh_lease(key) with a distributed lock or transaction shared by every worker. The lease must cover token reload, refresh, and save so concurrent workers cannot reuse a rotated refresh token. Supply the resulting user-bound AsyncAccessTokenProvider to the service. The OAuth token subject determines the personal memory account. The integration does not accept or send a user_id.
Complete interactive authorization during application bootstrap or account setup, not while processing a voice turn. Use one service and token context per authenticated participant. The application owns the token provider: close AsyncMemcodeOAuthClient and, when used, its DelegatingMemoryTokenProvider during connection or application cleanup.

​
Configuration

​
MemcodeMemoryService

​
AsyncMemcodeClient
default:"None"
Existing per-user asynchronous Memcode client. Supply either this or access_token_provider.
​
AsyncAccessTokenProvider
default:"None"
Async provider that resolves and refreshes a user-bound bearer token.
​
str
default:"\"https://memory.memcode.in\""
Memcode Memory API origin used when the service owns the SDK client.
​
str
default:"generated"
Stable, non-secret call or room ID used to derive retry-safe idempotency keys.
​
MemcodeMemoryConfig
default:"MemcodeMemoryConfig()"
Recall, context, ingest, and shutdown policy.
Important MemcodeMemoryConfig defaults are search_top_k=5, search_timeout_seconds=1.5, ingest_timeout_seconds=5.0, shutdown_timeout_seconds=2.0, and max_context_characters=4000.

​
Usage

Recall uses Memcode search_v2 with memory results only. Capture submits ingest_v2 with a deterministic idempotency key and waits only for the durable ingest receipt; Memcode’s asynchronous extraction does not block the pipeline. On EndFrame, pending writes receive a bounded flush window. On cancellation or cleanup, remaining owned tasks and clients are released idempotently.

​
Compatibility

This community package is maintained by Memcode outside Pipecat core. Consult its changelog for the exact latest tested Pipecat and memcode-sdk versions.