1
Your sources
Files, pages, recordings, notes — whatever your agent should remember.
Born working
Turn your files into a brain you can talk to.
Drop in your documents, recordings, and notes. Get back a private compartmentalized capsule you can query from Claude, ChatGPT, or Gemini — answering correctly the moment it finishes building.
Maple Harbor Guide
2 sources · 14 passages
Connect
…/c/maple-harbor/mcp
ClaudeChatGPTGemini
How it works
One pipeline, from a folder of files to a grounded answer.
A capsule is memory your agents read from and write to. The stack that builds it — the embeddings and the index — is disposable and rebuildable. The corpus, manifest, and eval set are the permanent asset.
2
Build
We distill the facts, embed them, and index them into a namespace that is yours alone.
Disposable — rebuildable at any time.
3
The capsule
corpusmanifesteval set
The permanent asset. Agents read and write — every write carries provenance and can be reverted.
4
Any agent
Ask, remember, recall — grounded answers with citations back to your sources.
Score each version against your own eval set, and export the whole capsule — rebuild the stack on any vector database, embedding model, or cloud.Read the open capsule spec →
Beyond RAG-as-a-Service
A capsule is memory your agents write to — not a read-only search box.
RAG-as-a-Service answers “search my docs” — a read-only pipeline welded to one provider. A Retenta capsule is memory an agent both reads from and writes to: every write is provenanced and revertible, every version is scored, and the corpus is portable across any vector database, embedding model, or cloud.
Write API
Agents read and write. Every write carries provenance — agent, session, source — and is revertible. Nothing is ever hard-deleted.
Versions you can score
Write an eval set once and grade any capsule version against it with an LLM judge, so you see the score before it ships.
Quality gates
Set a threshold. Below-gate versions are created but never promoted — bad memory can’t reach production.
Versioning
Content-addressed versions you can pin, diff, and revert. Production serves the version you pinned.
Portable export
Export the whole capsule — corpus, manifest, evals, write log — and rebuild it on any vector database, embedding model, or cloud.
MCP + REST per capsule
A first-class MCP server and REST API for every capsule, single-sourced from one tool spec so they never drift.
Connect in seconds
One call loads the whole memory spine.
The same tools over MCP or REST, single-sourced so they can’t drift. Recall the full state, append ordered progress, and remember findings — from Claude, ChatGPT, Gemini, or any headless harness.
recallrememberasksearchlog_progress
# Load the whole memory spine in one call
curl -H "Authorization: Bearer rk_live_…" \
…/c/maple-harbor/recall \
-d '{"query":"nightly triage"}'
# Write a fact back — provenance stamped, revertible
curl …/c/maple-harbor/remember \
-d '{"text":"CI is green tonight","topic":"ci-status"}'Pricing
Start free. The tier is the price — never a surprise bill.
We soft-warn at 80% of your quota and refuse new queries past 110% until the period resets or you move up a tier — never a silent bill past it. Downgrade or cancel and your data stays: every capsule, source and passage is kept, and you can export at any time.
Pro
$7/mo
3 capsules · 250k vectors
100,000 queries / mo
Everything in Developer, plus
- Write API + rollback
- Eval scores
Team
$29/mo
10 capsules · 1.5M vectors
500,000 queries / mo
Everything in Pro, plus
- Eval gates in CI
- Version pinning
Scale
$99/mo
30 capsules · 6M vectors
2,500,000 queries / mo
Everything in Team, plus
- Portability exports
- Priority support
FAQ
Questions, answered.
What can I put in a capsule?
PDF, Word, Excel, CSV, HTML, text and markdown, source code and notebooks, plus audio (.mp3, .m4a, .wav, .ogg, .webm, .flac) — audio is transcribed on build, then chunked and embedded like any other source. You can also point a capsule at a URL, or crawl a whole site from its sitemap. Images need OCR and PowerPoint isn’t supported yet, and a single very large file may not finish indexing.
Is my data isolated from other customers?
Yes. Every capsule is a private namespace, and tenant isolation is enforced in code — not just configuration — and covered by an automated isolation suite that runs on every change and must be green to merge. One customer can’t read, write, or list another’s capsules, vectors, storage, or usage.
Can I take my data with me, or am I locked in?
You can export the whole capsule — corpus, manifest, eval set, and write log — as a single archive with a re-materialize target, then rebuild it on any vector database, embedding model, or cloud. The embeddings and index are disposable build artifacts; the corpus and eval set are the permanent asset, and they’re yours.
Do agents only read, or can they write to memory?
Both. A capsule is memory an agent reads from and writes to. Every write carries provenance — agent, session, and source — and is revertible. Nothing is ever hard-deleted; a revert is a tombstone plus a log entry.
How do I know a new version of my capsule is any good?
Write an eval set — the questions your capsule has to get right — and grade any version against it with an LLM judge. Set a quality gate and below-threshold versions are still created but never promoted, so bad memory can’t reach production. The eval set travels with the capsule in every export.
Which agents and tools can connect?
Each capsule exposes a first-class MCP server and a REST API, single-sourced from one tool spec so they never drift. Connect from Claude, ChatGPT, Gemini, or any headless harness or cron loop.
What happens when I hit my plan’s limit?
We warn at 80% of your quota and refuse new queries past 110% until the period resets or you move up a tier — never a silent bill past it. Your data is never deleted over billing: downgrading or cancelling keeps every capsule, source and passage, and you can export at any time.
Encrypted, isolated, and yours to take with you.
AES-256 at rest and TLS 1.3 in transit, per-capsule isolation covered by an automated test, and a full export whenever you want it. We’re also honest about the one limit: to answer over your content we can read it, so we scan a narrow set for illegal material (CSAM and credible threats) and cooperate with law enforcement — never general monitoring, never selling your data.
How we protect your dataBuilt on Cloudflare.
Workers, D1, R2, Vectorize and Workers AI — one runtime, no cold-start theatre, no second cloud in the request path. Because the corpus, manifest and eval set are the permanent asset, the runtime underneath is an implementation detail we can change without touching what you own.
CloudflareRe-materialize targets — coming soon
AWS Bedrock
Google Vertex AIpgvectorYour capsule exports as corpus + manifest + eval set, so it can be rebuilt on infrastructure you choose. These targets are on the roadmap, not shipped.
More to come.
The stack is disposable. Make the memory permanent.
Embeddings and indexes are build artifacts you can rebuild anytime. The corpus, manifest, and eval set are the asset — and they’re yours to take anywhere.