Changelog
What's new
Release notes for the MemHQ memory API. New entries posted here first.
v0.22026-07-03Evidence-grounded answers
- Chunk-direct evidence lane: /ask now retrieves verbatim source dialogue alongside extracted memories and cites it. Answers can quote the exact detail your users said, not just the summarized fact.
- 83.2% on LoCoMo (up from 70.6%) — 8 points above Zep, 16 above Mem0's paper result. Per-question diff vs the prior configuration: 237 answers fixed, 24 regressed. Methodology and snapshot hash at /benchmark.
- Temporal anchoring fix: relative dates ("last week") now always resolve against the conversation date; pass metadata.occurredAt on /add for best results.
- Answer citations now include source-dialogue excerpts with session dates.
v0.12026-06-19Public launch
- Three-endpoint API — /add, /search, /ask — generally available.
- Per-user knowledge graphs with hybrid semantic + graph retrieval.
- Cited answers: /ask returns answer + source citations from memory.
- 70.6% on the LoCoMo long-term memory benchmark — above the current market leader (Mem0 at 67.1%). Full methodology published at /benchmark.
- TypeScript SDK (npm install @memhq/sdk) and Python SDK (pip install memhq).
- Remote MCP server at memhq.ai/mcp — plug MemHQ directly into any MCP-compatible agent.
- LangChain, Vercel AI SDK, and LlamaIndex adapters.
- Free tier: 1,000 adds, 5,000 searches, 500 asks per month. No credit card required.