VoiceMem vs Vectorize

The feature sets barely intersect even though both brands say 'memory.' Vectorize's differentiator is a learning loop: Retain stores, Recall retrieves, Reflect synthesizes patterns from tool-call failures and user corrections, and Mental Models apply that judgment in recurring situations. Its four networks exist to make an agent stop repeating mistakes, with per-user isolation and compound memory shared across collaborating agents. Recall is quoted under 100ms with parallel search, and it's model-agnostic, so you can swap LLMs without losing accumulated learning. Install is a single MCP skill command (npx add-skill), and it plugs into Claude Code, Cursor, Google ADK, and Slack.

VoiceMem is engineered for the audio pipeline, not text agents. It splits memory into a left brain (factual schemas/entities) and a right brain (persona, emotion, relationships), and runs a fully streaming path — segmentation, ASR, extraction, and graph writes while the user is still speaking. Speculative prefetching within a 0–300ms voice turn starts retrieval before the utterance finishes, with a published 134ms response time vs Mem0's 1,440ms. It claims 91.2% on LoCoMo at Top-5 (Mem0: 61.68%) and 69.44% on PersonaMem. Bundled modules include ASR, speaker verification, scene detection, emotion recognition, and local embeddings, plus a swappable memory engine and pluggable TTS. SessionBuffer purges temporary conversations at session end.

If your agent reads text and calls tools, Vectorize's reflexivity is the relevant capability. If your agent hears multi-party audio and must respond mid-turn, VoiceMem's streaming and emotion layers are the relevant ones — different problems, different feature lists.

Vectorize is freemium with two paths. The free path is a self-host under MIT: one Docker command, and you supply the Postgres and compute. The paid path is a cloud billed pay-as-you-go per token and per call, which suits teams who don't want to run infrastructure. The pricing risk Vectorize itself flags is high-retain-volume workloads sensitive to per-token memory costs — if you write a lot of memory, the meter moves. There's no seat-based pricing described, so cost scales with usage, not headcount.

VoiceMem is free under Apache-2.0 with no listed paid tier, managed service, or vendor support. That's genuinely zero license cost, but the operating cost is your time: model downloads, local warmup, source-level debugging, and self-hosted infrastructure. The vendor explicitly warns it's a v0.0.2 research project with no production SLAs. For a prototype or MVP with tight per-turn token budgets (it quotes ~430 memory tokens per query turn), that trade can be excellent value. For anything customer-facing on an uptime contract, the hidden cost is engineering hours and the absence of a support line when something breaks.

Bottom line: Vectorize gives you a paid escape hatch from ops; VoiceMem keeps everything free and hands the ops back to you. Budget accordingly — Vectorize's cost is a bill, VoiceMem's cost is an engineer.