Skip to content

compare.md

Latest commit

 

History

History
63 lines (47 loc) · 3.09 KB

compare.md

File metadata and controls

63 lines (47 loc) · 3.09 KB
Raw

Part of the ContextMemory docs. Back to README.

Why ContextMemory · how we compare

Why it exists

Problem ContextMemory
LLM forgets between messages Session markdown wiki + history inject
Memory is an opaque blob Wiki you can read, edit, version (asOf / supersede)
Docs live outside the chat Global Wiki via wiki_search in the agentic loop
Need tools without a second product Sandbox / MCP / HITL on the same /v1/chat/completions

How we compare

ContextMemory Mem0 Zep Letta
What it is Gateway + markdown wiki Memory library Temporal graph (cloud-first) Stateful agent runtime
You can open memory in an editor Yes No (API objects) Graph/UI Memory blocks / ADE
OpenAI chat drop-in Yes (you are the endpoint) No No Own agent API
Self-host complete AGPL, no gated core Partial / paid graph Graphiti + your DB Different product shape

vs Mem0: wiki + gateway that acts, not only a memory SDK.
vs Zep: bi-temporal markdown revisions, not Neo4j.
vs Letta: any OpenAI/MCP client; you do not adopt their agent OS.
LiteLLM is a router (complementary). LangChain is a client — point base_url here.

Why we are not RAG

We are an alternative when people search for RAG — not a RAG product. The harness is Cursor-style dynamic context discovery (digests + tools + artifacts), backed by a DB/wiki — without embeddings.

Classic RAG inject ContextMemory
Chunks stuffed into the prompt every turn Session wiki is budgeted memory; Global Wiki uses top-K digests + on-demand wiki_search / wiki_grep
Usually embeddings / vector DB Lexical / Postgres FTS (+ optional regex) on markdown you can open — deliberate, not a vector blob
Retrieval product bolted onto chat One agentic pipeline: wiki ± sandbox ± MCP ± HITL on the same /v1/chat/completions
Client often owns orchestration Client sends a normal chat body; the gateway runs the loop (compaction, lazy tools, subagents)

From the architecture docs: discovery tools in the loop, not a parallel RAG path — and classic RAG inject is explicitly N/A. Details: architecture-and-features.md.

Cloud vs self-host

Kortexio Cloud Self-host (this repo)
Ops None You run the .NET 9 gateway
Key cmk_live_… (no X-App-Id) cm_live_… + X-App-Id
Chat body Identical OpenAI /v1 Identical OpenAI /v1

Full guides: Cloud · Self-host / Docker

Architecture (30 seconds)

Client (Cursor MCP · OpenAI SDK · LangChain)
        │  memory_* tools  or  POST /v1/chat/completions
        ▼
ContextMemory — session wiki · Global Wiki · agentic loop · HITL
        │
        ▼
Ollama / vLLM / OpenAI / …     (+ optional sandbox / MCP integrations)

More detail: architecture-and-features.md