memoception
APIbenchmarksopen app →

Memory for Claude.

Memoception is an MCP connector: persistent, project-isolated memory your Claude can read and write. What Claude saves is what you see in the app — one memory, two doors.

Connect to Claude

Create an account in the app, then open Add to Claude to get your personal connector link (it carries a private token — treat it like a password).

claude.ai

  1. Copy your connector link from the app — it looks like https://www.engli.org/mcp?token=…
  2. In Claude: Settings → Connectors → Add custom connector → name it memoception → paste the link → Add.
  3. In any chat, enable memoception under the tools menu, then say “remember this”, “what do you know about …?”, “add this link to project X”, or “index this conversation into project X”.

Claude Code (terminal)

claude mcp add --transport http memoception https://www.engli.org/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

Perplexity

Perplexity doesn't take a one-click link (and doesn't support a token inside the server URL) — in Settings → Connectors → Add connector fill the form: Name memoception · MCP Server URL https://www.engli.org/mcp · Authentication API Key = your private connector key · Transport Streamable HTTP. The Connect to Perplexity button in the app pre-fills all of this with copy buttons.

Sanity-check any deployment (no token needed): curl https://www.engli.org/mcp returns the server card.

The ten tools

memory_storesave a fact, decision, preference, note, or todo into a project
memory_recallsemantic recall — one project, or your whole namespace
memory_ingest_filechunk + store a whole file (md, code, csv, json, jsonl, yaml, toml, ini, html, xml, sql, logs…)
memory_add_linkadd a web link — fetched & scraped into memory, and kept so Claude can call on it later
memory_index_historysave a whole conversation or project history in one call — incremental, re-runs skip known items
memory_projectslist your projects with memory counts and last activity
memory_forgetdelete memories by ids, project, or source file
memory_consolidatemerge near-duplicate memories into one stronger memory
memory_export_trainingturn a project's memory into prompt/completion or ChatML JSONL
grove_searchsearch the shared Engli Grove knowledge graph

What it remembers

Any format, any size. PDFs (scanned ones and images via OCR, English + French), Word/Excel/PowerPoint, audio & video (transcribed by a self-hosted Whisper), and 13 text formats — each cut along its natural seams; big files auto-split.
Semantic recall. Paraphrases match through concept expansion (“car” finds “automobile”); learned embeddings sharpen it further.
Project-isolated. Every memory is keyed by the namespace derived from your token — no other account can read, or even name, your data.

How recall works

Recall is a three-rung ladder, each rung degrading gracefully to the one below:

  1. Concept-expanded signatures. A conservative stemmer plus a curated synonym table map words to shared concepts, so “car” finds “automobile” with zero shared letters. Free, deterministic, runs anywhere.
  2. engli-embed-1024. In-process learned vectors served free from Memoception's own origin — no third-party call.
  3. Any external model. Point at any OpenAI-compatible embeddings endpoint when you want a specific model.

On the built-in paraphrase benchmark the ladder scores recall@1 of 1.0 versus 0.375 for the pre-ladder baseline — and that benchmark is a unit test, so quality can't silently regress. memory_consolidate periodically merges near-duplicates into one stronger memory.

Isolation & architecture

One account, one memory. Every door — a signed-in web session, a Bearer token, or the ?token= a claude.ai connector uses — resolves to the same namespace user:<id>. The web app, the Claude connector, and the CLI share exactly one memory; no other account can read, or even name, your data.
The brain runs on the Engli API (Next.js on Vercel, Neon Postgres over the serverless HTTP driver). Memoception is an isolated front end that talks to it cross-origin with your token; edge middleware handles CORS so a redirect can never break a request.
Extraction is client-side — files become text in your browser before anything is uploaded.

Built to last

The Engli TypeScript suite has 335 passing tests, including the recall benchmark and a full SSRF guard on link ingestion. The heavy runtime assets — the PDF worker, the OCR engine with English + French language data, and the Whisper runtime — are all self-hosted, so nothing depends on a third-party CDN. Its grove_search tool also bridges to the shared Engli Grove knowledge graph.

✦ Optimize

Inside any project, the ✦ optimize bar distills the whole project into its strongest form — duplicates merged, contradictions resolved (newest wins), rambling passages turned into precise facts, every name, number, and date preserved. The result lands as a new “project · optimized” box; re-running replaces it.
Write a prompt to steer it (or pick a preset from the ≡ menu: focus on new feature, security scan, refactor clean, upgrade version). Your first optimize is free; after that usage is metered per token against prepaid credits, and Pro runs a deeper multi-step optimization.

Storage & plans

Storage is auto-managed and isolated per account — or bring your own keys in Settings and your data sits in accounts you control. A storage meter under each project box (and inside it) shows exactly where you stand.
Free — 1 project · 100 GB. Pro ($39/mo) — unlimited projects · 1.5 TB per project · 9 TB total. Need more? Enterprise — book a call. Export everything at any time.

Get your connector link →