Changelog
What's new in MemorySync
Every release to the platform, from infrastructure upgrades to developer experience fixes.
v2.36
Memory for every agent stack: integrations, facts-only capture, migrations
Official packages connect MemorySync to the major agent frameworks, coding agents, voice stacks and automation platforms; every conversation-capture integration now stores curated facts instead of transcripts; and an account moves over from Mem0, Supermemory or Zep in one command.
integrationsplatformAPICLI- Framework packages: LangChain, LangGraph, Vercel AI SDK, CrewAI, Mastra, OpenAI Agents SDK, LlamaIndex, Google ADK, Pydantic AI, AutoGen, AG2, Microsoft Agent Framework, Strands, NVIDIA NeMo Agent Toolkit, Agno, CAMEL-AI, ChatDev and Eve, plus an AWS Bedrock reference architecture.
- Coding agents: one plugin package covers Claude Code and Cowork, Cursor, Codex, OpenCode, Devin and Windsurf, VS Code, OpenClaw, Hermes Agent and Antigravity, with a building-with-memorysync skill that teaches an agent the integration first try.
- Voice: LiveKit, Pipecat and ElevenLabs Agents, under a hard recall budget so memory never delays a reply.
- Automation and platforms: n8n nodes, a Zapier app, a Dify plugin and a Flowise node.
- Observability: AgentOps and Respan see every memory operation as a span through opentelemetry-instrumentation-memorysync; memory content is excluded by default.
- Facts-only capture: conversation turns are distilled into curated facts server-side, the same pipeline the connectors and the API use. Surfaces that replay history (LangChain chat history, the LangGraph store, OpenAI Agents sessions) stay byte-exact.
- Imports and migrations: background import jobs for large datasets, idempotent client_ref on every record, and one-command migration from Mem0, Supermemory and Zep.
- Accounts and CLIs: an agent can sign itself up and a person can claim the account later; the Node and Python CLIs are interchangeable and share stored keys.
- Fixes: v1 forget and summarize honour the API key’s environment; distilled facts no longer appear as extra transcript entries; replayed histories keep every turn.
v2.7
Model Context Protocol servers
Two hosted MCP servers let any MCP client read and write MemorySync memory, and search the documentation, without writing an integration.
MCPSDKsintegrations- Memory server at mcp.memorysync.io/mcp with 15 tools: search, add, list, read, update, delete, entity grouping, event history and ingestion status.
- Sign in from the client with OAuth 2.1 (PKCE, dynamic client registration); an API key still works for headless and CI use.
- Documentation server at docs.memorysync.io/mcp, unauthenticated, so a coding agent writes integration code from the current docs.
- Destructive tools preview by default and act only on an explicit confirmation. One-command setup for Claude Desktop, Claude Code, Cursor, Devin, VS Code, Codex and OpenCode.
v2.6
Retrieval quality and per-request timing
Ranking is stable across reworded queries, and every query response reports its own retrieval time.
RetrievalPerformance- Paraphrased questions produce the same ordering, through query expansion that falls back safely to the single-query path.
- Query responses include a latency field, separating server retrieval time from network and client overhead.
- Lower tail latency on large active memory sets; bulk writes report a per-item outcome.
v2.5
Continuous retrieval evaluation
Retrieval quality is measured on every change by a standing evaluation suite rather than assessed by hand.
BenchmarksRetrieval- Published the retrieval evaluation methodology: how relevance is scored and which query shapes are covered.
- LoCoMo and LongMemEval harnesses run against the same public API an integration uses.
- The suite runs in CI, so a ranking regression is caught before release; scoring is deterministic and comparable across releases.
v2.4
Unambiguous write outcomes
Writes report clearly whether a record was created, and deduplication no longer crosses environment boundaries.
IngestionDeduplication- Deduplication is scoped to the same project and environment, and fails open when similarity cannot be evaluated.
- Write responses distinguish stored, deduplicated and skipped outcomes.
- Re-summarizing the same set updates the existing summary; summary identifiers stay stable.
v2.2
Isolation and correctness under concurrency
A degraded upstream dependency for one organization no longer affects another, and usage counting is exact at the limit.
ReliabilityQuotas- Per-tenant failure isolation for model and embedding providers; recall degrades stage by stage instead of failing the request.
- Quota enforcement is an atomic check-and-increment: over-limit requests never count, and two concurrent requests cannot both take the last unit.
- Memory tier transitions are atomic; background jobs interrupted by a deploy resume instead of restarting.
v2.0
Throttling and correlation contract
Clients can base retry and correlation behavior on documented response data instead of inference.
APIObservabilityWebhooks- Rate-limit responses carry Retry-After with limit, remaining and reset headers; remaining quota is also returned on success.
- Every response returns a request id; inbound request and trace ids are honoured and searchable in the dashboard.
- Webhook retries use bounded exponential backoff, with a visible dead-letter state and per-endpoint delivery health.
v1.8
Enterprise identity and key rotation
Fewer manual steps to onboard an organization, and key rotation that needs no coordinated migration.
SSOSCIMAuditEncryption- Guided enterprise SSO setup covering domain verification, provider configuration and attribute mapping; SAML assertions are verified on every login.
- Directory provisioning updates accounts in place and deactivates on removal while keeping audit history.
- Encryption keys are versioned per record, so rotation keeps access to existing data without a migration.
- Audit entries carry a category and severity; SIEM forwarding gains a test action and per-destination status.
v1.6
API consistency and documentation rebuild
Consistent contracts across memory endpoints, and documentation reorganized around tasks rather than features.
APISDKsDocs- Field names, error shapes and pagination behave consistently across memory endpoints; structured errors carry a stable code.
- Request contracts match between the REST API and the Python and Node SDKs; project scope is resolved before any handler runs.
- Documentation rebuilt around task-based navigation, with exact request and response contracts per endpoint.
Changelog
Build on a platform that keeps shipping
Start on the free tier and get every improvement as it lands, with no migration work on your side.
Start free. No credit card required.