Agents are only as intelligent as the context they can reason over. GoodMem is the self-improving context layer for AI.

It learns from your documents and your agents’ searches, and builds private models that you own. Sovereign yet portable, it follows you to the cloud, your VPC, or a private server.

Compare deployment options ↓Contact sales

Replay of fine-tuning in GoodMem Cloud. Setup: pick a GoodMem instance and its Memory Spaces, choose the embedder, the reranker or both, see the estimated cost and time, and set a cost cap. The run reads the spaces, prepares training data, trains the models on a GPU and checks the results. Result of job 108: use the tuned model. Estimated change, plus 3.9 nDCG@10 points, 95% interval plus 2.8 to plus 5.0, compared with the same models untuned (GoodMem Embed Small and GoodMem Rerank Small) on 533 test searches held out from training and generated from the content. The tuned reranker, 1.7 GB, can be downloaded or deployed to a GPU at a private address. Scores out of 100: untuned embedder alone 81.8; with the untuned reranker 84.0; with the tuned reranker 87.8, the recommended setup; tuned embedder alone 86.2; tuned embedder with the untuned reranker 84.5; both tuned 88.2.

Over 533 searches, retrieval quality rose from 84.0 to 87.8 (nDCG@10) with an auto-tuned reranker.

Replay of fine-tuning in GoodMem Cloud. Setup and training were recorded on Nov 4, 2026 (job 104); the results are job 108, a later run on the same Memory Space with an improved trainer. Playback compresses recorded step times. See the fine-tuning benchmark on eight public tasks →

Retrieval Optimizer

Automatically selects the best retrieval stack for your data

The Retrieval Optimizer in GoodMem Cloud benchmarks embedder and reranker combinations — open-weight or API — on your real queries, then tells you which ones are statistically near-best. Evidence, not vibes.

  • Your queries, not a leaderboard Every candidate is scored on your own workload and held-out data.

  • Honest uncertainty 95% ranges and the odds that each pipeline is near-best, so you can trade a sliver of quality for cost or latency.

  • Keeps going As your data and traffic change, GoodMem re-tunes the stack — no labeling required.

cloud.goodmem.ai · Retrieval Optimizer
Result: Pipelines 1, 2, and 3 are strong choices. Pipelines 1, 2, and 3 each have at least a 95% chance of scoring within 0.05 NDCG of the best. Choose among them based on cost, latency, and operational fit.
  1. Embedder: Voyage 3 LargeReranker: Rerank 2.5 Mean NDCG 0.766 (95% range 0.712–0.812); 100% chance of being near-best, 99% chance of being best.
  2. Embedder: OpenAI Embedding 3 LargeReranker: Rerank 2.5 Mean NDCG 0.752 (95% range 0.694–0.801); 100% chance of being near-best, 1% chance of being best.
  3. Embedder: Voyage 3 LargeReranker: Qwen Qwen3 Reranker 8B Mean NDCG 0.729 (95% range 0.673–0.777); 100% chance of being near-best, 0% chance of being best.
  4. Embedder: OpenAI Embedding 3 LargeReranker: Qwen Qwen3 Reranker 8B Mean NDCG 0.719 (95% range 0.662–0.767); 67% chance of being near-best, 0% chance of being best.

Results from a real GoodMem Cloud optimization run.

Enterprise model training

Frontier-class AI you can own

Price-performance comparison for a custom BI agent over 20 Oracle EBS scenarios. Only the model changes.

Pass rate and cost per 1,000 conversations for four models on the BI-agent benchmark
ModelPass rateCost / 1K conv.
Claude Opus 4.8API only61.7%$2,380
PAIR-trained 35B model2×H100 you own55.9%$67
GLM-5.216×H10054.4%$260
GPT-OSS-120B1×H10030.0%$10

Frontier APIs are smart but expensive, and every call leaves your network. Small models can be self‑hosted but don’t perform well without tuning. For one customer’s BI agent, PAIR trained a 35B model that reached 91% of Opus 4.8’s score at 1/35 of the cost, running on two H100s.

Beyond text

Every page, ready for multimodal models

Real documents aren’t plain text. The answer is often in a table grid, a chart, or a drawing — exactly what text extraction flattens. GoodMem renders every page to an image at ingestion, even from Word, PowerPoint, and Excel, so a multimodal LLM can read the page the way a person would.

page image · 6-21
A catalog page titled Contractor's Grapple Matching Guide: a grid of excavator models against grapple models, where a dot means the grapple fits and a shaded cell means it does not.
  1. Query

    “Does a G136 grapple fit a 321D excavator?”

  2. GoodMem retrieves the page

    Extracted text

    321D  •  •

    Two dots — but under which grapples? The columns are gone.

    Page image

    The grid survives: the 321D row, and every column header above it.

  3. A multimodal LLM reads the page

    Answer: No.

    The 321D takes the G120 or G126 on a B linkage. Both G136 columns are shaded “No Match” on that row.

  • PDF
  • DOCX
  • PPTX
  • XLSX

Page images for PDF and all three Office formats, Excel included — rendered in-process, in pure Java, with no LibreOffice in the deployment.

Read how GoodMem captures pages

Two agents ask for the salary bands. Only one gets them.

Permissions are checked before anything is searched. If a key can’t read a space, the request fails outright, and nothing leaks from the spaces it can read.

Two agents send the same question. The support agent’s key is scoped to the Help Center, so its request for the Compensation space is refused with HTTP 403 before any search runs. The HR assistant’s key is scoped to Compensation, so it gets the L5 salary band.

“What’s the salary band for an L5 engineer?”

support-agent

scoped to Help Center

403

Permission denied

The key can’t read Compensation, so nothing was searched.

hr-assistant

scoped to Compensation

comp-bands-2026.xlsx · p. 3

Engineering salary bands
Engineering L4$148k – $176k
Engineering L5$182k – $214k
Engineering L6$221k – $262k
Stored in
Your PostgreSQL with pgvector. No proprietary store to migrate out of.
Runs in
GoodMem Cloud, your VPC, or fully air-gapped.
Logged
Per your policy: who asked, with which key, what they asked, and what came back.
Supply chain
Signed SLSA Build L3 provenance · distroless, non-root server image.
Audited
ISO/IEC 27001 · SOC 2 Type II.

Cloud or local. Your choice.

GoodMem keeps your pipeline portable

Use frontier APIs and the open model stack with the same GoodMem retrieval pipeline. Connect vLLM, TEI, Ollama, and OpenRouter alongside your hosted providers.

  • OpenAIHosted API
  • AWS
    AWS BedrockManaged models
  • OllamaLocal models

GoodMem

One retrieval pipeline

“The ability to swap providers, mix local and cloud infrastructure, and keep the retrieval pipeline consistent across all of them is genuinely impressive.”
B. — a GoodMem customer
Explore all integrations

Flexible deployment

Build agents on your own terms

  • Self-hosted

    Production-ready, free for commercial use.

    Free

    Self-host GoodMem

    What’s included:

    • Embed and ship anywhere, royalty‑free
    • Unlimited memories per node
    • gRPC API and SDKs in six languages
    • No Auto-Optimizer
  • Fully managed

    GoodMem Cloud

    Continuously optimized for your data, ready in seconds.

    From $15/mo

    Start free

    What’s included:

    • Auto-Optimizer fine-tunes your retrieval models
    • We run, monitor, scale, and update it
    • Usage-based pricing, no surprise overages
    • 14-day free trial, no credit card
  • Enterprise

    For OEMs, regulated industries, and service providers.

    Custom

    Contact sales

    What’s included:

    • On-prem or in your VPC
    • Auto-Optimizer on your own infrastructure
    • Custom LLM training by PAIR
    • Source code access and a 99.99% SLA
terminal

# Install goodmem

user@goodmem:~$ curl -s "https://get.goodmem.ai" | bash