Engrammic Context Service

Others store memories. We adjudicate claims.

---

What this is

Engrammic is a memory backend for AI agents that treats knowledge as something to be earned, not just stored. When an agent writes a fact, the system validates it, tracks where it came from, detects contradictions, and promotes it through epistemic layers only when corroboration warrants it.

The result is a knowledge graph where every node has a trust level, a provenance chain, and a lifecycle.

What this is not

• Not a vector database wrapper. Qdrant is one storage layer; the graph and epistemic machinery are the product.
• Not a RAG pipeline. Retrieval is one part of a larger write-time and read-time architecture.
• Not a scratchpad. Unverified observations decay. Claims require evidence. Beliefs require corroboration.
• Not drop-in compatible with systems that treat memory as a key-value store.

---

Architecture

Agents / Clients
       │
┌──────┴──────┐
│             │
MCP Server    FastAPI REST
(primary)     (admin)
│             │
└──────┬──────┘
       │
Service Layer
       │
┌──────┼──────┐
│      │      │
Memgraph  Qdrant  Redis
(graph)   (vector) (cache)

Surface	Purpose
MCP Server	Primary agent surface. Intent-based verbs (remember, learn, recall).
FastAPI REST	Admin operations (silo management, health, metrics).
SAGE	Background Dagster pipelines for synthesis, validation, and maintenance.

Full architecture: context/architecture.md

---

Quick start

Requires Docker Compose and uv.

git clone https://github.com/engrammic-ai/context-service
cd context-service
uv sync --all-extras

just up   # Start Memgraph, Qdrant, Redis
just dev  # Start the service

MCP server starts at http://localhost:8000/mcp.

Self-hosted deployment: docs/self-hosted/

---

Epistemic layers

Every node has a layer reflecting its trust level:

Layer	Type	How it forms
Memory	Observation	Agent writes via remember. Decays over time.
Knowledge	Claim	Agent writes via learn with evidence URI.
Knowledge	Fact	Promoted from Claim after corroboration (3+ sources).
Wisdom	Belief	Synthesized by SAGE from clustered Facts.
Intelligence	Commitment	Agent-declared or system-derived goals.

Write-gate

Claims require an evidence URI. The custodian validates citations and either promotes to Fact or rejects. No evidence = no write to knowledge layer.

Supersession

Updates create version chains, not overwrites. Old nodes stay with valid_to timestamps. Use trace(node_id) to walk provenance.

---

MCP tools

Tool	Purpose
remember	Store observation (Memory layer, no evidence required)
learn	Record claim with evidence (Knowledge layer)
recall	Retrieve by semantic query, node ID, or fusion
trace	Walk provenance chain
forget	Request deletion with optional cascade
tick	Lightweight engagement check
update	Supersede existing knowledge

Full reference: docs/api/mcp-tools-reference.md

---

Links

engrammic.ai	Website
docs.engrammic.ai	Documentation
research	Papers and benchmarks
primitives	Schema library
mcp	MCP proxy client

---

Contributing

1. just check must pass (mypy strict + ruff)
2. Depend on engine/protocols.py, not concrete stores
3. Knowledge-layer writes require evidence URIs
4. No commits directly to main

---

License

Apache 2.0. See LICENSE.