engrami 0.3.0

Auditable, compression-driven memory infrastructure for AI agents

Engrami

The transparent, verifiable memory layer for AI agents.

Graph + vector + relational memory in one self-hostable datastore — that your team can see, search, and trust. Plug it into any agent as an MCP server or call it from the SDK.

---

Why Engrami

Agents forget everything between sessions. Bolt-on memory tools (mem0, supermemory) fix recall — but they're black boxes: you can't see what's stored, why a fact changed, or what a change might break. For teams shipping real agents, that's a non-starter.

Engrami is built around a different promise: memory as a glass box.

	Engrami
One datastore	PostgreSQL + pgvector (vectors) + Apache AGE (native openCypher graph) + relational + a hash-chained journal. Neo4j optional. No separate vector DB.
Transparent	A live graph UI with search + filters; full provenance + journal timeline per fact; a plain-language change-log of every consolidation and forgetting decision.
Verifiable & deterministic	BLAKE3 hash-chained audit log, per-retrieval verification certificate, an input-hashed extraction cache (same input → same memory).
Active, not passive	A KG intelligence layer warns about cross-domain impacts at memory-time ("this fix touches a node on the checkout-latency path").
Self-hosted	Runs entirely in your VPC. Nothing leaves. Licensed via Cryptlex.

60-second quickstart

# 1. Start the stack (Postgres + pgvector + Apache AGE, and the API)
docker compose -f docker/docker-compose.yml up -d

# 2. Install the SDK/CLI
pip install -e .                 # or: pip install engrami

# 3. Initialize + store + recall
engrami init --tenant demo
engrami observe "Billing uses Postgres for ACID; the pricing call has a 200ms budget." \
  --tenant demo --source user_direct
engrami query "why postgres for billing?" --tenant demo

# 4. See it — the glass box
open http://localhost:8000/viz?tenant=demo

SDK:

from engrami import EngramiClient

async with EngramiClient(tenant="demo") as mem:
    await mem.add("We tried Redis price caching; it caused stale-price incidents.")
    ctx = await mem.search("can we cache prices?")
    print(ctx["context"])        # LLM-ready facts
    print(ctx["advisories"])     # proactive impact/contradiction warnings
    print(ctx["certificate"])    # audit receipt

MCP (shared in-cluster server):

engrami-mcp --transport http --port 8765      # agents connect to http://host:8765/mcp

What's inside

• Storage — GraphStore interface; default PostgresStore (pgvector + tsvector + recursive-CTE) or AgeStore (native openCypher via Apache AGE); Neo4jStore optional.
• Memory pipeline — security gate → multi-pass KG extraction (facts → entities/typed relations) → embeddings → bi-temporal nodes + edges, all journaled.
• Retrieval — hybrid vector + BM25 + RRF + multi-hop graph expansion, with a verification certificate and decompression-depth control.
• Compression — MERGE / ABSTRACT / PRUNE / ENCODE with pack guardrails, every action in the change-log.
• Intelligence — impact / dependency / contradiction advisories from the KG.
• Surfaces — CLI (engrami), MCP server (engrami-mcp, stdio or Streamable HTTP), HTTP API (engrami-api), Python SDK (EngramiClient).

Documentation

• MCP connector guide — connect an agent (HTTP or stdio), config, tools, the loop.
• Python SDK guide — install, quickstart, every method with examples + return shapes.
• Getting started (engineers) — install + an example for every CLI command, MCP tool, and SDK method.
• For executives — value, differentiators, compliance.
• Deploy in your VPC — step-by-step, with Helm + Terraform.
• Architecture · MCP tool reference · Domain packs

License

Engrami is proprietary software © Engrami LLP, licensed under the Engrami LLP Enterprise Software License Agreement (see LICENSE). It is licensed, not sold; production use requires a valid License Key issued via Cryptlex. Contact your Engrami account team for a Subscription. See DEPLOYMENT_VPC.md for activation in your VPC.