MemoryIntelligence API

Memory without meaning is just storage.

This API turns raw content (text, conversations, images) into structured meaning. Searchable, explainable, provenance-tracked. No black box. No hallucinations.

For Python backends: FastAPI, Django, Flask, Lambda, Cloud Functions.

---

What's new in 0.1.2

• ask() and list() are fixed. They previously returned empty results (hit.summary/hit.id as None, list() as []) because the client read the wrong response shape. Upgrade with pip install -U memoryintelligence.
• Now Apache-2.0 licensed (was a custom MIT-with-restrictions).

Full history in CHANGELOG.md.

---

Installation

pip install memoryintelligence

Get your API key: memoryintelligence.io/beta

---

Quick Start

from memoryintelligence import MemoryIntelligence

mi = MemoryIntelligence.from_env()          # reads MI_API_KEY

# Turn content into a structured, verifiable memory
note = mi.capture(
    "Discussed pricing with ACME. They prefer quarterly billing.",
    source="sales_call",
)

# Ask by meaning, not keywords
for hit in mi.ask("what did ACME say about billing?"):
    print(f"{hit.score:.2f}  {hit.summary}")

# Every memory has a receipt
if mi.verify(note.id):
    print("intact")

Upgrading from MemoryClient? The flat MemoryIntelligence client above is the recommended surface. The older MemoryClient (the mi.umo.* namespace, with multi-tenant user_ulid) still works for advanced use, and EdgeClient still serves regulated on-premise deployments, but MemoryClient is deprecated and will be removed in a future major. See Core Operations for the full eight-verb surface.

---

What It Does

• Processes → Extracts entities, topics, relationships, sentiment
• Searches → Returns ranked results with explainability
• Matches → Compares memories for relevance/similarity
• Explains → Shows why results matched (semantic, temporal, graph)
• Deletes → GDPR-compliant data removal with audit trail

No vector database setup. No chunking strategies. No embedding models to manage.

---

Security: API Keys Are Server-Only

Your API key grants full account access. Keep it server-side.

✅ Backend (environment variable):

mi = MemoryIntelligence.from_env()   # reads MI_API_KEY

❌ Never hardcoded:

# Don't - visible in logs/code
mi = MemoryIntelligence(api_key='mi_sk_beta_...')

For web apps: Use your API key in the backend, build your own endpoints with your auth, let your frontend call those.

---

What Makes This Different

Others	MemoryIntelligence
Store text, embed, hope	Extract meaning first
"Here are similar chunks"	"Here's why this matched"
No audit trail	Cryptographic provenance
Your data = their model training	We never train on your data

Setting the standard for memory.

---

Core Operations

Eight verbs, mapped one-to-one onto the public API:

note    = mi.capture(content, source="notes")   # POST   /v1/process
hits    = mi.ask(query, limit=5)                 # POST   /v1/memories/query
note    = mi.get(umo_id)                          # GET    /v1/memories/{id}
notes   = mi.list(limit=20)                       # GET    /v1/memories
proof   = mi.verify(umo_id)                        # GET    /v1/memories/{id}/proof
why     = mi.explain(umo_id)                        # GET    /v1/memories/{id}/explain
receipt = mi.forget(umo_id)                        # DELETE /v1/memories/{id}
notes   = mi.batch(["note one", "note two"])       # POST   /v1/batch
note    = mi.upload("meeting.mp3")                 # POST   /v1/upload

Results come back as small typed objects (Memory, Match, Proof, Receipt), each with a .raw escape hatch and item access, so note.id when it is clean and note["quality_score"] when you need a field we did not surface.

---

Async Support

An async twin of the flat client (AsyncMemoryIntelligence) is on the roadmap. Until then, the async surface is the older AsyncMemoryClient (the mi.umo.* namespace, deprecated alongside MemoryClient):

from memoryintelligence import AsyncMemoryClient

async with AsyncMemoryClient.from_env() as mi:
    results = await mi.umo.search(query, user_ulid="01ABC...")

---

Error Handling

from memoryintelligence import (
    MIError, AuthenticationError, NotFoundError, RateLimitError, ValidationError,
)

try:
    hits = mi.ask(query)
except AuthenticationError:
    ...                       # API key invalid or expired
except NotFoundError:
    ...                       # no memory with that id
except RateLimitError:
    ...                       # the client already retried with backoff
except ValidationError as e:
    ...                       # request shape was wrong
except MIError as e:
    print(e.status, e)        # any API error carries its status

---

Typed Everywhere

Full Pydantic models. Type hints on every method. No dict soup.

from memoryintelligence import Memory, Match, Proof

---

Examples & Docs

• Getting Started Guide — Full walkthrough
• FastAPI Integration — Complete example
• Advanced Usage — Batch processing, retry logic, custom configs

---

Support

• Docs: memoryintelligence.io/docs
• Issues: github.com/memoryintelligence/memoryintelligence-python-sdk/issues
• Beta: memoryintelligence.io/beta

---

MemoryIntelligence™ — Setting the standard for memory.

Built by somewhere