CognitiveXCOGNITIVEX · USE CASE
Memory for AI coding assistants
Stop re-explaining your codebase every session. One persistent memory across Cursor, Codex, Claude Code, and every other assistant: your decisions, conventions, and past fixes, recalled on demand.
WHY YOUR ASSISTANT FORGETS
Every new chat starts from zero.
AI coding assistants are stateless. Cursor, Codex, and Claude Code each load your repo, read a few files, and answer, then forget. The decision you explained yesterday (“we use Result types, never throw across the boundary”) is gone tomorrow. The bug you root-caused last week is rediscovered from scratch. You become the assistant’s memory, re-pasting the same context into every session.
Memory for AI coding assistants fixes this by giving the assistant a place to write down what it learns and read it back later. CognitiveX provides that layer: the Large Cognition Model (LCM). It exposes remember and recallover MCP, so any MCP-aware tool can persist and retrieve your project’s context. The memory is the model: it lives with you and your codebase, not inside one editor’s chat window.
FOUR TIERS OF PROJECT MEMORY
The context an assistant should never lose.
Semantic: facts
- Architecture decisions: “auth is JWT, sessions live in Postgres.”
- Conventions: error handling, naming, the lint rules you actually enforce.
- Stack constraints: “SSE not WebSockets behind the proxy.”
Episodic: events
- What you built last sprint and which files it touched.
- Bugs found and fixed, anchored to a date and a root cause.
- The reasoning behind a refactor you’d otherwise re-litigate.
Procedural: how-tos
- Your deploy ritual, test commands, and migration workflow.
- The footguns a new contributor (or a fresh chat) keeps hitting.
- Patterns the assistant should reach for in this repo.
A fourth tier, foundational, holds the durable identity of you and the project: goals, values, the non-negotiables that frame every decision.
THE DIFFERENCE
A stateless assistant vs. one with an LCM.
| Capability | Bare assistant | + Files / rules | + CognitiveX LCM |
|---|---|---|---|
| Survives a new chat session | No | Partial | |
| Shared across Cursor, Codex, Claude Code | No | Per-tool files | |
| Semantic recall (not keyword grep) | No | No | |
| Consolidates repeated context into conventions | No | No | |
| Decays stale notes by salience | No | Manual edit |
Static rules files (an AGENTS.mdor a cursor rule) are a good start, but they’re per-tool, they don’t do semantic recall, and you maintain them by hand. The LCM is shared, queryable, and consolidates on its own.
CONNECT IT
Three steps to a memory your tools share.
- Add the CognitiveX MCP server. Point your assistant at the CognitiveX MCP endpoint and add your API key. Cursor, Codex, and Claude Code all speak MCP. See the Cursor setup and the Codex setup for the exact config.
- Let it remember. Tell the assistant to call
rememberafter decisions, bug fixes, and milestones, or wire it into your project rules so it does so automatically. Storing memories is free; you’re only billed on recall. - Recall on demand. At the start of a task, the assistant calls
recallwith a query (“how does auth work here?”) and gets back the relevant decisions, events, and how-tos, no matter which tool wrote them.
The same memory is reachable from the HTTP API and the cogx SDK, so CI jobs, scripts, and your own agents can read and write it too. Full reference lives in the docs.
QUESTIONS
FAQ
Does it work across different coding tools?
Yes. Recall and remember are exposed over MCP and HTTP, so a decision made in Cursor is recallable in Claude Code and vice versa. The memory belongs to you and your project, not to any single editor.
Is this just a vector store?
No. A vector store retrieves what you stored. The LCM also consolidates, promoting recurring episodes into durable conventions, decaying stale notes by salience, and detecting patterns during overnight dream consolidation, so the memory reflects how you actually work.
What does it cost?
Storing memories is always free. You spend recall credits only when you read, priced by depth (foundational, standard, deep). The free tier includes 100 recall credits a month; see pricing for the full tiers.
Give your assistant a memory that lasts.
Wire CognitiveX into the tools you already use, or try the cognition layer in the iCog app first.