engrammemory-ai 0.1.2

Memory infrastructure for AI agents. Your Qdrant, your hardware, our intelligence.

Memory infrastructure for AI agents. Your Qdrant, your hardware, our intelligence.

npm · Dashboard · API Docs · Community Edition

pip install engrammemory-ai

Quick Start

from engrammemory import Engram

client = Engram(
    api_key="eng_live_xxx",
    qdrant_url="http://localhost:6333"
)

# Store — embedded & deduplicated by Engram, stored in YOUR Qdrant
client.store("User prefers TypeScript and dark mode", category="preference")

# Search — three-tier recall (hot → hash → vector)
results = client.search("What does the user prefer?")
for r in results:
    print(f"[{r.tier}] {r.content} ({r.score:.2f})")

# Forget
client.forget("mem_abc123")

Why Engram?

	Engram	Mem0	Supermemory	Zep
Data ownership	Your Qdrant, your hardware	Their cloud	Their cloud	Their cloud
Three-tier recall	Hot → Hash → Vector	Single-tier	Single-tier	Two-tier
Stateless intelligence	Embeds + classifies, never stores	Stores everything	Stores everything	Stores everything
Edge/fleet support	Per-device agents, zone isolation	No	No	Limited
Local-first + overflow	Store locally, overflow to cloud	Cloud only	Cloud only	Cloud only
Open community edition	OpenClaw skill, full recall engine	No	Partial	No

Three-Tier Recall

Every search flows through three tiers automatically:

results = client.search("database config")
for r in results:
    print(f"[{r.tier}] {r.content} ({r.score:.2f})")
    # [hot]    PostgreSQL 15 with read replicas (0.94)   — sub-ms, cached
    # [hash]   Migrated from MongoDB last quarter (0.81)  — O(1) LSH lookup
    # [vector] Old DB credentials in vault (0.67)         — full ANN fallback

Auto-Recall for Agent Prompts

# Inject relevant memories into your agent's system prompt
memories = client.recall("User is asking about their database setup")
system_prompt = f"You know: {', '.join(m.content for m in memories)}"

Async

from engrammemory import AsyncEngram

async with AsyncEngram(api_key="eng_live_xxx") as client:
    await client.store("User prefers TypeScript")
    results = await client.search("language preferences")

Multi-Agent / Fleet

# Each agent gets its own namespace
client = Engram(api_key="eng_live_xxx", project="icu-floor-3")

# Store with agent tracking
client.store("Patient allergic to penicillin", category="fact", agent="tablet-icu-3a")

# Search scoped to project
results = client.search("allergies", agent="tablet-icu-3a")

MiroFish Integration

Replace Zep with Engram in your MiroFish .env:

MEMORY_PROVIDER=engram
ENGRAM_API_KEY=eng_live_xxx
ENGRAM_QDRANT_URL=http://localhost:6333

API Reference

Constructor

Engram(
    api_key: str,                          # Your Engram API key
    qdrant_url: str = "http://localhost:6333",  # Your Qdrant instance
    base_url: str = "https://api.engrammemory.ai",
    collection: str = "agent-memory",
    project: str | None = None,            # Memory isolation namespace
    max_retries: int = 3,
    timeout: float = 30.0,
)

Methods

Method	Description
store(content, category?, importance?, metadata?, agent?)	Store a memory (embedded + deduplicated + compressed)
search(query, limit?, min_score?, category?, agent?)	Three-tier recall: hot → hash → vector
recall(context, top_k?, agent?)	Auto-inject relevant memories for agent prompts
get(memory_id)	Get a specific memory by ID
forget(memory_id)	Remove from all three tiers
forget_by_query(query)	Forget memories matching a search
list(limit?, offset?, category?, agent?)	List memories with filtering
health()	Check Engram + Qdrant connectivity

Models

from engrammemory import Memory, SearchResult, StoreResult, HealthStatus

Model	Fields
Memory	id, content, category, metadata, created_at, access_count
SearchResult	memory, score, tier ("hot", "hash", "vector")
StoreResult	id, stored, deduplicated
HealthStatus	status, qdrant_connected, version

Exceptions

from engrammemory import EngramError, AuthenticationError, RateLimitError, NotFoundError, ValidationError

Exception	HTTP Code	When
AuthenticationError	401	Invalid or missing API key
RateLimitError	429	Too many requests (auto-retries with Retry-After)
NotFoundError	404	Memory ID doesn't exist
ValidationError	400/422	Invalid request body
EngramError	Other	Base class for all errors

Community Edition

Don't need cloud? The community edition runs the full three-tier recall engine locally as an OpenClaw skill with no API dependency.

Links

• Node SDK — npm install engrammemory-ai
• Dashboard
• API Docs
• GitHub
• Community Edition