G2

Raafeh S.

AI Engineer

Small-Business (50 or fewer emp.)

8/11/2026

"Dependable Hybrid Memory Retrieval That Makes “Remember This User” a System Feature"

4/5

What do you like best about Mnemo?

The biggest win for me is reliable long-term memory retrieval for agents specifically hybrid search that finds the right user fact even when the memory store is noisy.

Concrete example: I bulk-loaded ~500 memories for a test user (mostly filler notes) and planted 6 real facts rice preference, Costa Rica packing details, timezone, reply style, a lamp bulb change, and vendor evaluation notes. Across 3 search rounds, every needle came back at rank 1. That is the feature I care about most: I can store atomic facts once, then ask natural questions later (“how should I respond?”, “what bulb did I put in the bedside lamp?”) and get the right memory without hand-tuning retrieval.

How it improved the workflow

Instead of stuffing growing chat history into every prompt, I can keep a scoped memory store per user (scope: { type: "user", id: "..." }) and pull only what’s relevant before generating a reply.

Batch create (items[]) made load testing and seeding profiles fast — hundreds of memories in ~20–25 seconds in batches of 50.

Content deduplication was an unexpected benefit: re-posting the same fact didn’t create junk duplicates, which keeps the store cleaner for agent use.

Performance

Warm hybrid search felt production-usable (API total often ~20–25ms; client round-trip ~300ms). First queries after a big ingest were slower, then settled quickly.

AI / Intelligence

The hybrid pipeline (vector + lexical + fact strategies in my runs) is what makes this more than “just embeddings.” Intent classification + ranked results meant preference/style queries and factual queries both worked without me building a custom RAG stack.

Integrations / API

I mainly used the HTTP API. The OpenAPI surface is clear once you find it, and the mental model is simple: create memories → search → inject into the agent prompt. That’s the loop I’d use regularly in a real product.

Pricing / ROI (early view)

ROI for me is engineering time saved: not building user-scoped memory, indexing, dedup, and retrieval ranking myself. If you’re shipping a multi-session copilot, that infrastructure is usually weeks of work; Mnemo compresses it to an API call.

UI / UX & onboarding (honest note)

I didn’t live in the dashboard UI day-to-day value came from the API. Onboarding via the Python SDK/docs was the weaker part (SDK lagged the live API). Once on the correct API contract, the product experience was strong.

Bottom line: What I like best is that Mnemo makes “remember this user” a dependable system feature, not a prompt hack — and in my tests, the retrieval quality held up under load. Review collected by and hosted on G2.com.

What do you dislike about Mnemo?

The main downside is onboarding friction from docs/SDK lagging the live API not the core memory/search quality.

Specific pain

I installed getmnemo==0.1.2 and followed examples like from getmnemo import Mnemo and memory.memories.create(...). That surface isn’t in the published package (import is mnemo, and there’s no memories.create).

The shipped client still sends the old body (content / actorId / query). Live API 0.2.0 expects items[], scope, and q, so first create calls returned 400 until I switched to OpenAPI directly.

Naming drift adds confusion (getmnemo vs mnemo, older LedgerMem/Proofly leftovers, key prefixes). Fine once you know it; rough for a first-hour evaluation.

Workflow impact

Instead of “install → 10-line quickstart → value,” I spent time reverse-engineering the real contract. For a SaaS memory product, that first-run path matters as much as retrieval quality.

Other rough edges

Invalid memoryType (e.g. profile) can fail an entire batch, not just the bad item and allowed values weren’t obvious from the SDK.

After bulk ingest, the first searches were colder; then latency settled. Manageable, but worth documenting.

I didn’t get much day-to-day value from UI/UX because the SDK path pushed me straight to raw HTTP.

What I don’t dislike

Once on the correct API, create/search/dedup/scoped recall worked well. My dislike is mostly developer experience and version alignment, not the underlying memory engine.

Ask for the team: ship an SDK + docs that match production the same day the API changes, with a working copy-paste quickstart and a short migration note from 0.1 → 0.2. Review collected by and hosted on G2.com.

What problems is Mnemo solving and how is that benefiting you?

Problem 1: Agents forget users between sessions

LLMs are stateless. Without an external memory layer, a support/sales/copilot agent loses preferences, constraints, and prior facts every new chat.

Benefit: I can store durable facts per user (e.g. “prefers terse technical replies,” timezone, packing details) and retrieve them later, so replies stay consistent without re-asking.

Problem 2: Prompt stuffing doesn’t scale

Dumping full history into context gets expensive, noisy, and still misses the one fact you need.

Benefit: Scoped hybrid search lets me pull a small set of relevant memories before generation. In my tests, planted facts still ranked #1 inside a ~500-memory haystack so “find the right memory” worked under load, not just in a toy demo.

Problem 3: Building memory infra yourself is slow

User-scoped storage, dedup, indexing, and retrieval ranking is real engineering work if you DIY with a vector DB + custom pipeline.

Benefit: Mnemo compressed that to an API loop: create memories → search → inject into the agent prompt. Batch writes and content dedup saved setup time during seeding/load tests.

Problem 4: Multi-user isolation

In a real product, User A’s preferences must not bleed into User B.

Benefit: scope (e.g. user:stress_user) gives a clean tenancy model I’d use in production multi-tenant agents.

Net benefit for me

Mnemo turns “remember this user” from a prompt hack into a system feature. That improves agent quality (more personal, fewer repeated questions) and reduces the infra I’d otherwise build and maintain myself especially valuable when evaluating or shipping AI products that need continuity across sessions. Review collected by and hosted on G2.com.